Achieving Non-Inclusive Cache Performance with Inclusive Caches Temporal Locality Aware (TLA) Cache Management Policies

Size: px
Start display at page:

Download "Achieving Non-Inclusive Cache Performance with Inclusive Caches Temporal Locality Aware (TLA) Cache Management Policies"

Transcription

1 Ahieving Non-Inlusive Che Performne with Inlusive Ches Temporl Lolity Awre (TLA) Che Mngement Poliies Amer Jleel, Eri Borh, Mlini Bhndru, Simon C. Steely Jr., nd Joel Emer Intel Corportion, VSSAD Msshusetts Institute of Tehnology (MIT) {mer.jleel, eri.borh, mlini.k.bhndru, simon..steely.jr, Abstrt Inlusive hes re ommonly used by proessors to simplify he oherene. However, the trde-off hs been lower performne ompred to non-inlusive nd exlusive hes. Contrry to onventionl wisdom, we show tht the limited performne of inlusive hes is mostly due to inlusion vitims lines tht re evited from the ore hes to stisfy the inlusion property nd not the redued he pity of the hierrhy due to the duplition of dt. These inlusion vitims re inorretly hosen for replement beuse the lst-level he (LLC) is unwre of the temporl lolity of lines in the ore hes. We propose Temporl Lolity Awre (TLA) he mngement poliies to llow n inlusive LLC to be wre of the temporl lolity of lines in the ore hes. We propose three TLA poliies: Temporl Lolity Hints (TLH), Erly Core Invlidtion (ECI), nd Query Bsed Seletion (QBS). All three poliies improve inlusive he performne without requiring ny dditionl hrdwre strutures. In ft, QBS performs similr to non-inlusive he hierrhy. Keywords inlusion; non-inlusion; exlusion; replement I. INTRODUCTION As CMPs beome widespred nd the gp between proessor nd memory speeds ontinues to widen, it is impertive tht proessor rhitets design n effiient nd high performing he hierrhy. One of the key design hoies for multilevel he hierrhy is whether or not to enfore inlusion [6, 27, 7]. While inlusion gretly simplifies the he oherene protool [9, 6], it limits performne when the size of the lrgest he is not signifintly lrger thn the sum of the smller hes. In suh senrios, CPU rhitets resort to non-inlusive [27] or exlusive [7] he hierrhies. This pper fouses on improving inlusive he performne without srifiing its benefits. The inlusion property requires tht the ontents of ll the smller hes of multi-level he hierrhy be subset of the lst-level he (LLC) [6]. When line is evited from the LLC, inlusion is enfored by removing tht line from ll the hes in the hierrhy. We refer to he lines invlidted in the smll hes s result of inlusion s inlusion vitims. The smll hes, from here on referred to s ore hes, hide the temporl lolity from the LLC when they servie requests. Sine replement stte is only updted to Most Reently Used (MRU) on he hits, the LLC replement stte of hot lines onstntly servied by the ore hes deys to lest reently used (LRU) in the LLC. As result, the hot lines beome ndidtes for evition in the LLC. The number of inlusion vitims drmtilly inreses when multiple pplitions ompete for the LLC or when the LLC is not signifintly lrger thn the sum of ll the ore hes. A stright forwrd mehnism to eliminte inlusion vitims would be to remove the requirement tht ore hes be subset of the LLC. Suh he, known s noninlusive he [0], llows he lines to reside in the ore he(s) without lso being duplited in the LLC. In doing so, non-inlusion inreses the effetive pity of the he hierrhy [27, 29]. Unfortuntely, non-inlusion elimintes the nturl snoop filter benefit tht n inlusive LLC provides, thus breking the oherene benefits tht ome with inlusivity [6]. While snoop filters [5, 22, 23] n be used in ddition to the LLC, suh strutures inrese the hrdwre overhed [29] nd verifition omplexity [9]. It would be idel to design he hierrhy tht redues (if not elimintes) the frequeny of inlusion vitims while providing the oherene benefits of inlusion. Thus, our gol is to bridge the performne gp between inlusion nd non-inlusion by improving the mngement of n inlusive LLC. It is widely held belief tht the primry benefits of noninlusive he hierrhy ome from the inrese in the effetive hing pity [, 27, 29]. However, we show tht the first order benefit of non-inlusion is the elimintion of inlusion vitims nd not the extr he pity. In light of this observtion, we illustrte tht inlusive hierrhies n perform similr to non-inlusive hierrhies by preserving hot lines in the ore hes nd extending the life time of these lines in the LLC. We propose Temporl Lolity Awre (TLA) inlusive he mngement poliies to redue the frequeny of hrmful inlusion vitims. The suess of TLA poliies requires identifying lines tht hve high temporl lolity in the ore hes nd preventing LLC replement of these lines until their temporl lolity is exhusted in the ore hes. We propose three TLA he mngement poliies where the identity of hot lines is either onveyed to, derived by or inferred by the LLC. Temporl Lolity Hints (TLH) onvey the temporl lolity of hot lines in the ore hes by sending hints to the LLC to updte its replement stte. With the sme temporl informtion s the ore hes, the LLC is signifintly less likely to hoose vitim tht will use n inlusion vitim. Erly Core Invlidtion (ECI) derives the temporl lolity of line before it beomes LRU in the LLC. The LLC hooses line erly nd invlidtes tht line in the ore hes, while leving the line in the LLC. By observing the ore s subsequent request, the LLC derives the temporl lolity.

2 Core Request Core Request Core Request Miss Flow Fill Flow L LLC vitim evit bk-invlidte Miss Flow Fill Flow L LLC vitim Miss Flow Fill Flow L LLC vitim vitim fill Memory Memory Memory () Inlusive Hierrhy (b) Non-Inlusive Hierrhy () Exlusive Hierrhy Figure : Summry of Che Hierrhies. Query Bsed Seletion (QBS) infers the temporl lolity of line in the LLC by querying the ore hes. The LLC selets line for replement nd queries the ore hes for pprovl. The LLC uses the informtion provided by the ore hes to infer the temporl lolity nd selet vitim. QBS employs the heuristi tht the ore hs better informtion bout temporl lolity, nd it is best suited to propose whih lines n be evited. Our evlutions on 2-ore, 4-ore, nd 8-ore CMPs show tht the three TLA he mngement proposls improve inlusive he performne with QBS performing similr to non-inlusive he. We lso show tht the proposed poliies do not require ny dditionl hrdwre strutures nd only require extr messges in the system. The orgniztion of the pper is s follows: Setion II provides more in depth motivtion for the work; Setion III introdues TLA he mngement; Setion IV desribes our experimentl methodology, followed by our experimentl results in Setion V. Setion VI disusses relted work nd we summrize in Setion VII. II. MOTIVATION Figure illustrtes the different types of he hierrhies. An inlusive he hierrhy requires tht the ontents of the smller he levels be subset of the LLC. When line is evited from the LLC, inlusion is enfored by bkinvlidting (i.e., removing) tht line from the smller he levels (if present). The pity of n inlusive he hierrhy is equl to the size of the LLC. Inlusive hes re benefiil beuse they serve s nturl snoop filter. When n LLC lookup results in miss, no snoops need to be sent to the ore hes beuse they re gurnteed to not be present in the ore hes. A non-inlusive he hierrhy, on the other hnd, does not gurntee tht the smller he levels be subset of the LLC. As suh, the pity of non-inlusive he hierrhy rnges between the size of the LLC nd the sum of ll levels in the he hierrhy. Finlly, n exlusive he hierrhy requires tht the ontents of ny he level re not replited in ny other he level. Exlusive hierrhies first fill he lines t the smllest level in the hierrhy nd fill them into subsequent lrger he levels upon evition from the smller he levels. Exlusive hes require higher bndwidth sine even len vitims from the ore hes hve to be written to the LLC. The pity of n exlusive he hierrhy is equl to the sum of the sizes of ll the hes in the hierrhy. Unfortuntely, non-inlusive nd exlusive hes inrese the pity of hierrhy by srifiing the nturl snoop filter benefits of inlusive hes. In generl, the hoie of he hierrhy is dependent on the rtio of the size of smll hes (i.e., ore hes) to the LLC. Figure 2 ompres the performne of non-inlusive nd exlusive LLCs to n inlusive LLC for vrious ore he to LLC size rtios. The figure shows tht inlusive hes typilly perform worse thn non-inlusive nd exlusive hes. When the LLC size is more thn 8X lrger thn the ore hes, the performne of non-inlusive nd exlusive LLCs re similr to n inlusive LLC. However, when the LLC shrinks, both non-inlusive nd exlusive LLCs perform signifintly better thn inlusive LLCs. Modern dy Intel miroproessors (like the Intel Core i7 proessor) hve :8 he rtio nd implement inlusive LLCs [4] while AMD miroproessors (like the Phenom II proessor) hve :4 he rtio nd implement non-inlusive LLCs []. The figure shows tht inlusive LLCs with :8 rtio hve n verge of 3% (mx 2%) worse performne thn non-inlusive/exlusive LLCs. Similrly, inlusive LLCs with :4 rtio hve n verge of 8% (up to 33%) worse performne thn non-inlusive/exlusive LLCs. For suh he rtios, it is better not to enfore inlusion. Both noninlusive/exlusive LLCs improve performne over inlusive LLCs by eliminting inlusion vitims nd inresing the effetive hing pity of the hierrhy. Sine we nnot hnge the requirement tht inlusion duplite lines in the Performne Reltive to Inlusion Exlusive LLC : 2 : 4 : 8 : 6 Rtio of Core Che Size to LLC Size Figure 2: Performne of Non-Inlusive nd Exlusive LLCs.

3 Next Ref () Bseline Inlusive Che (b) Temporl Lolity Hints (TLH) () Erly Core Invlidtion (ECI) (d) Query Bsed Seletion (QBS) (t = 0) b I L: Hit! L2: - b I L: Hit! --- L2: LRU updte b I L: Hit! L2: - b I L: Hit! L2: - d (t = ) b I L: reple - L2: reple I b I L: reple - L2: reple I b I L: reple - L2: reple I L2: ECI b I L: reple - L2: reple I d d d I d (t = 2) d b L: Hit! L2: - d b L: Hit! --- L2: LRU updte d b L: reple I - L2: Hit! d b L: Hit! L2: - e (t = 3) d d b L: reple d - L2: reple d d b L: reple d - L2: reple b d d b L: reple d - L2: reple b L2: ECI d d b L: reple d - L2: reple b move to MRU (t = 4) e d e d b L: reple d - L2: reple b e e d L: Hit! --- L2: LRU updte Figure 3: Illustrtion of Temporl Lolity Awre Che Mngement Poliies. e e d L: Hit! L2: - e e d L: Hit! L2: - LLC, we fous on eliminting inlusion vitims to improve inlusive he performne. Inlusion vitims mnifest on CMPs whenever pplitions with working set sizes tht fit in the ore hes onurrently exeute with pplitions tht hve working set sizes tht re greter thn or equl to the size of the LLC. In suh situtions, the ore hes hide ll the temporl lolity of the ore he fitting (CCF) pplitions from the LLC. Consequently, the LLC replement stte of he lines belonging to the CCF pplitions dey towrds LRU when LLC fitting (LLCF) or LLC thrshing (LLCT) pplitions exerise the LLC. As result, the CCF pplitions suffer from inlusion vitims even though they hve high temporl lolity in the ore hes. The problem of inlusion vitims is not intrinsi to CMPs lone. It n lso mnifest in single-ore, single-threded system where non-unified instrution nd dt hes shre unified LLC. For exmple, streming pplition whose instrution working set fits in its instrution he n suffer from inlusion vitims when referenes to non-temporl dt evit the instrution working set from the LLC. Mintining inlusion nd voiding inlusion vitims for ode lines is pplible for rhitetures tht require ode lines to be oherent due to self-modifying ode (e.g. x86). Inlusive he performne n be improved if the LLC knew of the temporl lolity of lines in the ore hes. To ddress this problem, the next setion proposes Temporl Lolity Awre he mngement poliies for LLCs. III. TEMPORAL LOCALITY AWARE (TLA) MANAGEMENT In multi-level hierrhy, the ore hes t s filter nd hide the temporl lolity from the LLC. Thus, the LLC only observes the referenes tht miss in ll levels prior to the LLC. For lines tht hve high temporl lolity in the ore hes, the replement stte of the line in the LLC n dey to LRU. The line eventully gets evited from the LLC nd due to the inlusive hierrhy, the line is lso evited from the ore hes (n inlusion vitim). To illustrte this problem, onsider the following unfiltered he referene pttern:..., b,,,, d,, e,, f,,.... Note tht the referenes to '' hve high temporl lolity. Figure 3 illustrtes two-level inlusive he hierrhy with fully ssoitive 2-entry L he nd fully-ssoitive 4- entry LLC. Both the L nd LLC strt with ll entries invlid. The boxes nd rrows represent the LRU hin, with the MRU line on the left nd LRU line on the very right. The figure strts fter the referene to. Figure 3 illustrtes the behvior of n inlusive he hierrhy. The figure shows tht in L '' is the MRU entry nd '' is the LRU entry. In the LLC, '' is the MRU entry, nd the LRU entry is n invlid line, I. The next referene to '' hits in the L he nd updtes '' to the MRU position. Sine '' hits in the L he, only the L replement stte is updted nd the LLC replement stte stys unhnged. The next referene, 'd', misses in the hierrhy nd reples lines '' nd 'I' in the L nd LLC respetively. Line '' is now the LRU line in both the L nd LLC.

4 L hit (updte LRU) TLH (updte LRU) L LLC b I L LLC b I Vitim Bk Invlidte Erly Core Invlidte L LLC b I Yes! No! Evit it Updte LLC LRU, Choose next LLC vitim, repet query Cn LLC Evit This Line? MRU Next Vitim LRU MRU Figure 4: Temporl Lolity Awre (TLA) Che Mngement Poliies. Next Vitim () Temporl Lolity Hints (TLH) (b) Erly Core Invlidtion (ECI) () Query Bsed Seletion (QBS) LRU MRU Next Vitim LRU The next referene to '' hits in the L he. Agin, '' is moved to the MRU position of the L he nd no replement updtes our in the LLC. The next referene 'e' misses in both the L nd LLC. Sine '' is now the LRU entry in the LLC, it is hosen for evition, even though it is the MRU line in the L he. Due to the enforement of inlusion, line '' is n inlusion vitim nd is removed from the L he leving n invlid entry behind (not shown in the figure). Line 'e' tkes the ple of '' in the LLC nd the invlid entry left behind by '' in the L he. Line 'e' is promoted to the MRU position in both the hes. The next referene, '', misses in both the L nd LLC. For the bove referene pttern, line '' misses to memory on every ninth referene to the L he, despite its high temporl lolity. If the LLC were non-inlusive or if the temporl lolity of '' ws not hidden from the LLC by the L, '' would never hve been bk-invlidted nd suffer ore he miss. Reduing the hrmful effets of inlusion requires tht lines with high temporl lolity in the ore hes not be hosen for evition in the LLC until their temporl lolity in the ore hes is exhusted. This implies tht the LLC replement stte should be mde wre of the temporl lolity of lines in the ore hes. A. Temporl Lolity Hints (TLH) Inlusion vitims n be eliminted by onveying the temporl lolity of lines in the ore hes to the LLC. A diret mehnism would be to inform the LLC of the unfiltered temporl lolity. On hit in the ore hes, temporl lolity informtion n be onveyed to the LLC by issuing Temporl Lolity Hint (TLH) (see Figure 4). The hint is non-dt request tht updtes the replement stte of the line in the LLC. The heuristi used is tht if line is hot in ore he, it is worth preserving in the LLC. We illustrte the use of TLHs in Figure 3b. At the L he, the replement behvior remins the sme s the bse se. For every L he hit, the line is moved to the MRU position in the L he nd TLH is sent to the LLC. The LLC uses the hint to promote tht line to the MRU position in its LRU hin. Thus, in our exmple referene pttern, L he hits on '' updte the LRU stte in both the L nd LLC. Thus, when the referene to 'e' ours, '' is not in the LRU position nd is not evited from the LLC. By onveying the temporl lolity of, TLHs prevent inlusion vitims. TLHs n signifintly redue the number of inlusion vitims beuse they keep the LLC replement stte up-todte with the ore hes. However, the downside is tht they send request to the LLC for every hit in the ore hes. Optimiztions n be mde by filtering the number of TLHs sent from the ore hes to the LLC. For exmple, in multilevel hierrhy, the seond level he n issue TLHs insted of the first level he. Alterntively, the L he n issue TLHs for non-mru lines. However, even with filtering of TLHs to the LLC, the number of requests to the LLC is extremely lrge nd does not sle well with inresing number of ores. Thus, for the purpose of this study, we evlute the use of TLHs s limit study to determine the potentil performne of voiding inlusion vitims ltogether. Insted of using TLHs, we propose lterntive low overhed solutions: Erly Core Invlidtion (ECI) nd Query Bsed Seletion (QBS). B. Erly Core Invlidtion (ECI) Erly Core Invlidtion (ECI) is method for deriving the temporl lolity of line. The intent is to invlidte line from the ore hes erly (but retin the line in the LLC) nd observe ore's response to derive the line s temporl lolity. If the line is hot in ore he, the ore will re-request it. If this request rrives before the line gets evited from the LLC, it will be hit. As with ll hits, the temporl lolity t the LLC uses replement stte updte. If ore does not rereferene the line, or the re-referene tkes too muh time, the line will be evited by the next miss to tht set. Sending n ECI to the ore hes is reltively stright forwrd. As illustrted in Figure 4b, upon miss in the LLC, the stndrd vitim flow hppens normlly vitim is seleted nd evited from the LLC nd ore hes to mke room for the inoming line. ECI dds nother step to the miss pth by seleting nother vitim (the next LRU line) nd sending n ECI to the ore hes to hve tht line evited. The line still remins in the LLC 2 though. We illustrte the behvior of ECI in Figure 3. Unlike TLHs no updte trffi to the LLC is required. When the referene to 'd' misses in both the L nd LLC, the miss t the LLC is hndled s usul by eviting the LRU line, I. With ECI, the next LRU line, '', is lso evited from the L he (it is still retined in the LLC). This results in 'd' being MRU in L nd LLC, nd 'I' nd '' being LRU in the L nd LLC. Like the Core i7, diretory is mintined with eh LLC line to determine the ores to whih bk-invlidte must be sent [3]. 2. The diretory bits of the LLC line re updted s usul.

5 respetively. The next referene to '' (t t=3), misses in L but hits in the LLC. This updtes the replement stte of '' in the LLC. By premturely eviting line '' from the ore hes nd then observing the subsequent request for, the LLC derived tht '' ws hot, nd ptured the temporl lolity of ''. Note, however, tht ECI is time sensitive. The ore must re-referene the line within window of opportunity. For instne, hd new line been referened t t=3 before the rereferene to, '' would hve been evited. A line hs to be hot reltive to the time window, time window tht will vry depending on the level of ontention t the LLC. With ECI, the number of lines tht need to be invlidted in the ore hes on eh LLC miss n be either one or two. If n erly invlidted line hs not been re-referened, it is not in ny ore hes. Thus, when the next miss ours, tht line is hosen s the vitim, nd beuse it is not in ny ore hes, the bk-invlidte request is not required. Only the ECI needs to be sent out for the next LLC vitim. In ontrst, if n erly invlidted line is re-referened by ore before LLC evition, its LLC replement stte is updted. Thus, the next miss fter suh hot line resue will need to find vitim to evit, whih will inur norml invlidte s required by inlusion long with the ECI to derive the temporl lolity of the next potentil vitim. Note tht ECI hppens in the shdow of the miss to memory, nd thus is not time ritil. As ompred to TLHs, ECI is lower trffi solution to derive line s temporl lolity. This is beuse the number of TLHs is diretly proportionl to the number of ore he hits (whih is high) while ECI request trffi is diretly proportionl to the number of LLC misses (whih is orders of mgnitude smller thn the number of ore he hits). There re trde-offs though. The erly invlidte is predition tht the line will not be needed by the ore hes. If the predition is inorret, the sheme llows time window (until the next miss to tht set) for the ore hes to orret the predition by re-requesting the line. If the ore re-requests the line in tht time window, then the downside is tht wht would've been hit in the ore hes hs beome hit in the LLC. The dded lteny of getting the dt from the LLC insted of the ore hes is trded ginst extending the line's lifetime in the LLC. The more signifint downside is when the re-request for the line misses the time window. The onsequene is n LLC miss tht must be servied from min memory. The time window for ore to re-referene n erly invlidted line restrits the bility of the LLC to fully derive the temporl lolity, nd thus limits the potentil benefits of ECI. C. Query Bsed Seletion (QBS) To ddress the drwbks of ECI, we propose Query Bsed Seletion (QBS). Like ECI, QBS is lso initited from the LLC. Insted of invlidting line erly nd hving the ores re-request it, QBS voids invlidting lines in the ore hes ltogether. QBS queries the ore hes nd hs them indite if line is resident in the ore hes nd uses tht informtion to infer the temporl lolity of those lines. Lines resident in the ore hes re inferred to hve high temporl lolity nd re not evited from the LLC. As illustrted in Figure 4, on n LLC miss, the he ontroller selets the LRU line s the vitim nd queries the ore hes to determine if the line n be evited from the LLC. If the line is not present in ny of the ore hes, QBS evits this line to mke spe for the inoming line. However, if the urrent vitim line is present in ny of the ore hes, QBS updtes the line's LLC replement stte to MRU nd extend its lifetime in the LLC. Sine vitim still needs to be seleted, new vitim is hosen nd the proess repets. The QBS vitim seletion proess is typilly hidden by memory lteny. Should the dt rrive from memory before QBS selets vitim, the he ontroller n either wit for QBS to finish seleting vitim or fore QBS to selet the next vitim. Alterntively, the he ontroller n limit the number of queries issued on n LLC miss. When the mximum is rehed, the next vitim line is seleted for replement nd no further queries re sent to the ore hes. We illustrte the behvior of QBS in Figure 3d. When the referene to 'e' misses in the LLC, QBS queries the ore hes to determine whether it n evit the first potentil LLC vitim, ''. When the ore hes re queried for '', the ore responds tht the line is present in the L he. Consequently, the LLC replement stte of is updted to MRU nd query for the next vitim, 'b', is sent. Sine b is not resident in the L he, the ore llows 'b' to be repled from the LLC. Thus, when '' is re-referened, it hits in the L he. QBS hs prevented '' from being n inlusion vitim. By preventing the LLC from eviting lines tht re still resident in the ore hes, QBS prevents hot lines from beoming inlusion vitims. QBS ddresses the time window problem of ECI nd voids LLC misses. IV. EXPERIMENTAL METHODOLOGY A. Simultor We use CMP$im [6], Pin [9] bsed tre-driven x86 simultor for our performne studies. Our bseline system is 2-ore CMP. Eh ore in the CMP is 4-wy out-of-order proessor with 28-entry reorder buffer nd three level he hierrhy. Without loss of generlity, we ssume single-thred per ore. The L nd L2 hes re privte to eh ore. The L instrution nd dt hes re 4-wy 32KB eh while the L2 he is unified 8-wy 256KB. The L nd L2 he sizes re kept onstnt in our study. We support two L red ports nd one L write port on the dt he. The bseline lst-level he (LLC) is unified 6-wy 2MB he tht is shred by both the ores in the CMP. We ssume bnked LLC with s mny bnks s there re ores in the system. Like the Intel Core i7, we model non-inlusive L2 he nd only the lst-level he enfores inlusion 3. All hes in the hierrhy use 64B line size. For replement deisions, the L nd L2 hes use the LRU replement poliy while the LLC uses the Not Reently Used (NRU) replement poliy 4 [2, 4]. We model strem prefether tht trins on L2 he misses nd prefethes lines into the L2 he. The prefether hs 6 strem detetors. The lod-to-use 3. Modern proessors [, 2, 3] use non-inlusive L2 hes. If the L2 were inlusive, TLA poliies n be pplied t the L2 he to perform similr to non-inlusive hes. We verified this in our studies. 4. The inlusion problem is independent of the LLC replement poliy. The problem ours with LRU replement s well s more intelligent replement poliies (e.g. RRIP [4]). We verified this in our studies.

6 TABLE I. MPKI of Representtive SPEC CPU2006 Applitions In the Absene of Prefething st bzi l de gob h26 hmm lib mf per pov sje sph wrf xl L MPKI (64KB) L2 MPKI (256KB) LLC MPKI (2MB) ltenies for the L, L2, nd LLC re, 0, nd 24 yles respetively. We model 50 yle penlty to min memory nd support 32 outstnding misses to memory. The he hierrhy orgniztion nd ltenies re bsed on the Intel Core i7 proessor [3]. The proposed poliies do not rely on the speifi ltenies used. We hve verified tht the proposed poliies perform well for different ltenies inluding pure funtionl he simultion. We ompre the poliies using both the throughput 5 nd he performne metris. We model n interonnet with fixed verge lteny. Bndwidth onto the interonnet is modeled using fixed number of MSHRs. Contention for the MSHRs models the inrese in lteny due to dditionl trffi introdued into the system. All trnstions, inluding those for ECI nd QBS, use existing dt pths nd ontend for the MSHRs. We do not model bndwidth limittions for the Temporl Lolity Hints (TLH). TLHs serve s limit study on potentil performne improvement if temporl lolity informtion from ll levels in the hierrhy were vilble t the LLC. We lso ompre the performne of the proposed TLA poliies to non-inlusive nd exlusive he hierrhies. Che ltenies re identil to the bseline inlusive he hierrhy for both these he hierrhies. A non-inlusive he hierrhy is modeled by not sending bk-invlidtes to the ore hes when repling lines from the LLC. To model n exlusive he hierrhy we modify both the hit nd miss pths. Lines re invlidted in the LLC upon he hits. As for the miss pth, new lines re inserted into the ore hes first. These lines re inserted into the LLC only fter they re evited from the ore hes. Exlusive hes typilly require higher LLC bndwidth beuse even len vitims from the ore hes must be inserted into the LLC. We do not model this inresed bndwidth, nd thus our results for exlusive hes re optimisti. B. Benhmrks The SPEC CPU2006 benhmrks were first grouped into three different tegories bsed on their L, L2, nd LLC he hit behvior. The first tegory is those pplitions whose working set fits in the ore hes. We refer to these pplitions s ore he fitting (CCF) pplitions. The seond tegory is those pplitions whose working set fits in the LLC. We refer to these pplitions s LLC fitting (LLCF) pplitions. Finlly, the third tegory is those pplitions whose working set is lrger thn the LLC. We refer to these pplitions s LLC thrshing (LLCT) pplitions. Of ll the SPEC CPU2006 benhmrks, we seleted five from eh 5. We ompred the performne of the TLA poliies on both the weighted speedup nd hmen-firness metris. Sine the TLA poliies do not introdue ny firness issues, they perform similr to the throughput metri for both weighted speedup nd hmen-firness metris. tegory to over the spetrum of hit/miss behvior in the different levels of the he hierrhy. The 5 representtive SPEC CPU2006 benhmrks were ompiled using the i ompiler with full optimiztion flgs. Representtive regions for the SPEC benhmrks were ll olleted using PinPoints [20]. Tble I lists the 5 SPEC CPU2006 benhmrks nd their misses per 000 instrutions (MPKI) in the L, L2, nd LLC when run in isoltion. The MPKI numbers re reported in the bsene of prefether. Bsed on the MPKI vlues in Tble I, delii, h264ref, perlbenh, povry, nd sjeng ll hve smll miss rtes in the L2 he. This implies tht these benhmrks hve working sets tht fit into the ore hes. Thus, for our bseline he hierrhy, we n lssify these pplitions s CCF pplitions. The benhmrks gobmk, libquntum, mf, sphinx3, nd wrf ll hve working set sizes tht re lrger thn the bseline LLC size. This is beuse these pplitions hve lmost s mny misses s there re esses to the LLC. Thus, for our bseline he hierrhy, these pplitions n be lssified s LLCT pplitions. Finlly, the benhmrks str, bzip2, lulix, hmmer, nd xlnbmk ll hve working set sizes lrger thn the L2 he but benefit from the LLC. Thus, for our bseline he hierrhy, we n lssify these pplitions s LLCF pplitions. In generl, CCF pplitions re vulnerble to inlusion relted evitions when onurrently exeuting with LLCT or LLCF pplitions. To study the effets of inlusion, we rn ll possible two-threded ombintions of the 5 SPEC benhmrks, i.e. 5 hoose 2 05 worklods. To provide insights on when TLA poliies re benefiil, we seleted 2 worklod mixes (listed in Tble II) to showse results. However, we provide results for ll 05 worklods. We simulted 250 million instrutions for eh benhmrk. Simultions ontinue to exeute until ll benhmrks in the worklod mix exeute t lest 250 million instrutions. If fster thred finishes its 250M instrutions, it ontinues to exeute to ompete for he resoures. We only ollet sttistis for the first 250 million instrutions ommitted by eh pplition. This methodology is similr to existing work on shred he mngement [5, 2, 25]. TABLE II. Worklod Mixes Nme Apps Ctegory Nme Apps Ctegory MIX_00 bzi, wrf LLCF, LLCT MIX_06 hmm, xl LLCF, LLCF MIX_0 de, pov CCF, CCF MIX_07 de, wrf CCF, LLCT MIX_02 l, gob LLCF, LLCT MIX_08 bzi, sje LLCF, CCF MIX_03 h26, per CCF, CCF MIX_09 pov, mf CCF, LLCT MIX_04 gob, mf LLCT, LLCT MIX_0 lib, sje LLCT, CCF MIX_05 h26, gob CCF, LLCT MIX_ st, pov LLCF, CCF

7 V. RESULTS AND ANALYSIS A. Temporl Lolity Hints A Limit Study Figure 5 shows the throughput normlized to the bseline inlusive he hierrhy when ll requests t different levels in the hierrhy send TLHs. As reminder, we do not model ny bndwidth onstrints when sending TLHs to the LLC. The first three brs in the figure show the performne when the L instrution he lone sends TLHs (TLH-IL), the L dt he lone sends TLHs (TLH-DL), nd when both L instrution nd dt hes send TLHs (TLH-L). The x-xis represents the different worklod mixes nd the br lbeled All represents the geomen of ll 05 worklods used in the study. TLH-IL nd TLH-DL lone improve throughput by 5% or more for the worklod mixes MIX_08, MIX_09, MIX_0, nd MIX_. For these worklod mixes, the performne improvements re dditive when both the IL nd DL hes send TLHs (i.e., TLH-L). This is beuse these worklod mixes re omposed of CCF pplition tht is running onurrently with n LLCT or LLCF pplition. Sending TLHs from the ore hes prevents LLCT nd LLCF pplitions from eviting the he lines of CCF pplitions from the LLC. This llows TLHs to improve the performne of n inlusive he by s muh s 3%. The figure lso shows tht homogeneous worklod mixes of CCF pplitions (MIX_0 nd MIX_03) or worklod mixes tht hve ombintion of LLCT nd LLCF pplitions (worklod mixes MIX_00 nd MIX_02) reeive no benefits from TLHs. For suh mixes, inlusion vitims re not problem beuse either ll benhmrks hve working sets tht fit in the ore hes or ll hve working set sizes tht re lrger thn the ore hes. On verge, TLH-IL, TLH-DL, nd TLH-L improve throughput by 2.6%, 2.5% nd 5.2% respetively. Figure 5 lso presents the performne of TLHs when only the L2 he hits send TLH to the LLC (TLH-L2). On verge, TLH-L2 hs roughly two-thirds the performne of TLH-L. For exmple, MIX_0 observes only 7% improvement in throughput with L2 generted TLHs ompred to 24% improvement from L generted TLHs. MIX_0, onsists of benhmrks libquntum nd sjeng. From Tble I, we see tht libquntum hs no lolity in ny of the hes while sjeng hs good L he lolity. Thus, TLH-L2 is unble to refresh the LRU stte of the lines belonging to sjeng in the LLC. Figure 5 lso illustrtes tht the performne of TLH-L nd TLH-L2 is not dditive when both the L nd L2 hes send TLHs (TLH-L-L2). In ft, TLH-L provides the bulk of the performne improvements when ll levels in the hierrhy send TLHs. Figure 5 shows tht TLH-L bridges 85% of the gp between inlusive nd non-inlusive he hierrhies while TLH-L2 only bridges 45% of the gp between inlusive nd non-inlusive hes. Figure 5 lso shows the s-urve for the throughput improvement of TLH- L, TLH-L2, nd non-inlusive LLC for ll 05 two-ore worklod mixes. The s-urve is sorted bsed on non-inlusive he performne. In the figure, x represents TLH-L2 nd tringles represent TLH-L. The figure shows tht TLH-L losely trks the performne of non-inlusion for the bulk of the worklods. These results show tht the performne of non-inlusive he hierrhy n be hieved by llowing n inlusive LLC to be wre of the globl temporl lolity of lines tht re resident in the LLC. While TLH-L bridges the gp between inlusive nd non-inlusive hes, the number of LLC requests inrese by lmost 600X. In omprison, TLH-L2 inreses the number of requests to the LLC by bout 8X. We onduted sensitivity study on the frtion of hits tht n send hints to the LLC. When %, 2%, 0%, nd 20% of ore he hits in the L he generte TLHs, they bridge the gp between inlusive nd non-inlusive he hierrhies by 50%, 60%, 75%, nd 80% respetively. However, even when % of hits in the L he send TLHs, the number of LLC requests still inreses by 6X or more. Thus, while TLH-L n potentilly bridge the performne gp between inlusive nd non-inlusive he hierrhies, the high bndwidth demnds required by TLHs require n lternte low bndwidth tehnique to llow the LLC to derive or infer the globl temporl lolity. B. Erly Core Invlidtion Figure 6 presents the performne of ECI for the different worklod mixes. The figure shows tht ECI improves performne by more thn 7% for four out of the 2 worklods. The four worklods orrespond to the sme worklods tht benefit from TLHs, i.e. worklods where CCF pplition is onurrently exeuting with LLCT or Performne Reltive to Inlusion.3.2 TLH-IL TLH-DL TLH-L TLH-L2 TLH-L-L2 Performne Reltive to Inlusion.3.2 TLH-L TLH-L2 MIX_00 MIX_0 MIX_02 MIX_03 MIX_04 MIX_05 MIX_06 MIX_07 MIX_08 MIX_09 MIX_0 MIX_ All (05) Figure 5: Performne of Temporl Lolity Hints (TLHs). Worklod ID

8 Performne Reltive to Inlusion.3.2 ECI Performne Reltive to Inlusion.3.2 ECI MIX_00 MIX_0 MIX_02 MIX_03 MIX_04 MIX_05 MIX_06 MIX_07 MIX_08 MIX_09 MIX_0 MIX_ All (05) Figure 6: Performne of Erly Core Invlidtion (ECI). LLCF pplitions. On verge, ECI bridges 55% of the gp between inlusive nd non-inlusive hes. To illustrte this, Figure 6 lso presents the s-urve for ECI for ll 05 worklods. Aross ll worklods, ECI provides up to 30% improvement in performne with the worst se outlier losing only.6% performne. Unlike TLHs, ECIs re initited from the LLC to derive the temporl lolity of lines. The dditionl bk-invlidte requests required by ECI on verge is less thn 50% (in the worst se it doubles). While the inrese in bk-invlidtes might seem signifint, the extr messges introdued by ECI re reltive to the number of LLC misses. On verge, our studies show n LLC miss rte of 5 misses per 000 yles. ECI inreses the number of bk-invlidtes from 5 trnstions per 000 yles to 7 trnstions per 000 yles. Sine the number of trnstions is smll to begin with, the inrese in bk-invlidte requests from the LLC is negligible nd n esily be sustined by the ores. As our studies show, the inrese in bk snoop trffi from the LLC to the ores does not impt performne. The first order benefit of ECI is from voiding misses to min memory (Che misses re disussed in Setion D). C. Query Bsed Seletion Figure 7 shows the performne of QBS when pplied to different hes in the hierrhy. QBS-IL nd QBS-DL ensure tht he lines resident in the IL or DL hes re not evited by the LLC. QBS-L ensures tht he lines tht re resident in ny L he re not evited by the LLC. QBS- Worklod ID L2 ensures tht he lines resident only in the L2 re not evited by the LLC. Finlly, QBS-L-L2 ensures tht he lines resident in ny he in the hierrhy re not evited by the LLC. The figure shows tht QBS-IL onsistently performs better thn QBS-DL. Sine instrution he misses stll the front-end of the pipeline from supplying instrutions to the bk-end, on verge, it is more importnt to retin ode lines in the LLC until their temporl lolity is exhusted in the instrution he. On verge, QBS-IL improves performne by 2.7% nd QBS-DL improves performne by.6%. QBS-L is dditive of QBS-IL nd QBS-DL nd improves performne by 4.5%. Figure 7 lso shows tht QBS-L onsistently performs better thn QBS- L2 QBS-L2 improves performne on verge by.2%. Agin, this is beuse the L hes filter out the temporl lolity from the L2 hes. Finlly, Figure 7 shows tht QBS- L-L2 outperforms non-inlusion on verge. To illustrte this, Figure 7 lso presents the s-urve for QBS for ll 05 worklods. On verge, QBS improves performne by 6.5% while non-inlusion improves performne by 6.%. We lso evluted versions of QBS tht limit the number of queries tht the LLC n mke before seleting vitim. We evluted query limits of, 2, 4, nd 8 for QBS nd observed 6.2%, 6.5%, 6.6% nd 6.6% performne improvements respetively. The bseline inlusive he hierrhy n be thought of hving 0 queries. Performne does not signifintly improve beyond two queries beuse the totl pity of the on-hip hes is equl to the pity Performne Reltive to Inlusion.3 QBS-IL QBS-DL QBS-L QBS-L2 QBS-L-L2.2 MIX_00 MIX_0 MIX_02 MIX_03 Figure 7: Performne of Query Bsed Seletion (QBS). MIX_04 MIX_05 MIX_06 MIX_07 MIX_08 MIX_09 MIX_0 MIX_ All (05) Performne Reltive to Inlusion.3.2 QBS Worklod ID

9 % Fewer LLC Misses Reltive to Inlusion MIX_00 TLH-L TLH-L2 ECI QBS Exlusive LLC MIX_0 MIX_02 Figure 8: Che Performne Reltive to Inlusion. MIX_03 MIX_04 of two wys of the LLC. On verge, sending only one query is suffiient to hieve the bulk of the performne benefits. When QBS only sends one query, like ECI, our studies showed 50% inrese in bk-invlidte requests from the LLC to the ore hes. As we disussed for ECI, the 50% inrese in bk-invlidte requests is negligible beuse the trffi is smll to begin with nd the inrese in bkinvlidte requests from the LLC to the ores n be esily sustined. Our studies show tht the inrese in trffi from the LLC to the ores does not impt performne. Like ECI, the first order benefit of QBS is from voiding LLC misses. D. Effet of TLA Mehnisms on Che Misses Sine inlusion vitims require re-fething hot lines from memory, we use LLC miss redution s the metri to mesure the goodness of the proposed TLA poliies. While redution in L he misses n lso be ompred, they re not pplible to ECI or L2 TLH beuse they both evit lines from the L he. However, they both void requests to memory by preserving hot lines in the LLC. Thus, we present redution in LLC misses sine this metri is pplible to ll the TLA poliies. We verified tht the TLA proposls redue L he misses where pplible. Figure 8 presents the redution in LLC misses for the different TLA poliies, non-inlusive he, nd exlusive he when ompred to the bseline inlusive he. The br lbeled All represents the verge redution in LLC misses for ll 05 worklod mixes. On verge, L TLH redues LLC MIX_05 MIX_06 MIX_07 MIX_08 MIX_09 MIX_0 MIX_ ALL(05) % Fewer LLC Misses Reltive to Inlusion QBS Worklod ID misses by 8.2%, L2 TLH by 4.8%, ECI by 6.5%, QBS by 9.6%, non-inlusive hes by 9.3%, nd exlusive hes by 8.2%. On verge, QBS performs better thn ECI beuse QBS elimintes the time window problem ssoited with ECI by updting the replement stte instntly insted of witing for the line to be resued by the ore 6. This llows QBS to perform similr to non-inlusive hes without the omplexity of non-inlusion. Figure 8 lso presents n s-urve tht ompres redution in he misses ross ll 05 worklod mixes for QBS. Like non-inlusion, QBS redues he misses by s muh s 80%. These results gin emphsize tht non-inlusive hes primrily ddress inlusion vitims nd not extr pity. Only exlusive hes tke dvntge of the extr he pity. E. Summry of All TLA Poliies Figure 9 summrizes the performne of the TLA mehnisms proposed in this pper ompred to the bseline inlusive he. The figure shows tht QBS performs similr to non-inlusive he hierrhy. To ensure tht the performne benefits of the TLA mehnisms re only from reduing inlusion vitims, we lso evluted the performne 6. A version of QBS tht bk-invlidtes lines from the ore hes but updtes the LLC replement ws lso evluted. This modified QBS version is similr to ECI in tht lines re evited from the ore hes. We find tht the modified QBS poliy performs similr to the proposed QBS mehnism. This implies tht the bulk of QBS benefits re from voiding memory lteny (not LLC hit penlty). Performne Reltive to Inlusion.3.2 TLH-L TLH-L2 ECI QBS Performne Reltive to Non-Inlusion.3.2 TLH-L TLH-L2 ECI QBS Exlusive LLC MIX_00 MIX_0 MIX_02 MIX_03 MIX_04 MIX_05 MIX_06 MIX_07 MIX_08 MIX_09 MIX_0 MIX_ All (05) MIX_00 MIX_0 MIX_02 MIX_03 MIX_04 MIX_05 MIX_06 MIX_07 MIX_08 MIX_09 MIX_0 MIX_ geomen () Bseline Inlusive Che (b) Bseline Non-Inlusive Che Figure 9: Performne of TLA Che Mngement Poliies In the Presene nd Absene of Inlusion.

10 Performne Reltive to Inlusion TLH-L TLH-L2 ECI QBS Exlusive LLC :2 :4 :8 :6 Rtio of Core Che Size to LLC Size Figure 0: Slbility of TLA Poliies to Different Che Rtios. impt of these mehnisms on non-inlusive he (where inlusion vitims do not exist). If there were some other benefits besides reduing inlusion vitims, we would expet benefits from TLA mehnisms on non-inlusive he. Figure 9b summrizes the performne of the TLA mehnisms normlized to the performne of non-inlusive LLC. The results show tht TLA he mngement poliies improve the performne of non-inlusive he by only 0.4.2%. These results show tht the TLA mehnisms provide insignifint improvement in performne for noninlusive he. However, they signifintly improve inlusive he performne. Thus, we onlude tht the first order benefits of non-inlusive he re primrily from voiding inlusion vitims. Finlly, the figure shows tht on verge, n exlusive he hierrhy hs 2.5% better performne thn non-inlusive he hierrhy. This shows tht eliminting inlusion vitims n signifintly bridge the performne gp between inlusive nd exlusive he hierrhies. One inlusion vitims re eliminted, the remining gp between inlusive nd exlusive hes is primrily due to the extr pity in the hierrhy. F. Slbility to Different Che Rtios Figure 0 presents the slbility of the proposed TLA mehnisms for different ore he to LLC rtios. The figure shows the performne of the TLA mehnisms for MB, 2MB, 4MB, nd 8MB LLC. The L2 to LLC rtios for these he sizes re :2, :4, :8, nd :6. In generl, reduing the LLC size while keeping the ore hes onstnt requires better LLC mngement. When the LLC is not signifintly lrger thn the ore hes, exlusive or non-inlusive hes signifintly improve performne ompred to the bseline inlusive hes. Both TLH-L nd QBS signifintly bridge the gp between inlusive nd non-inlusive hes. We find tht TLH-L does not perform s well s QBS for the :2 rtio beuse hot lines servied by the L2 he were suffering from inlusion vitims. We found tht TLH-L-L2 performs similr to QBS for this onfigurtion. Unlike TLHs (whih re imprtil due to the required mount of bndwidth), we find tht low bndwidth solution suh s QBS mthes noninlusive he performne for the different he rtios. G. Slbility to Different Number of Cores Figure presents the performne of QBS when the totl number of ores in the CMP re inresed. We reted Performne Reltive to Inlusion QBS (4-ore CMP) QBS (8-ore CMP) Worklod ID Figure : QBS Performne with Inresing Core Count. ore nd 8-ore worklod mixes nd evluted QBS performne on CMP where the rtio of the ore he to LLC is mintined t :4. The 4-ore CMP hs 4MB LLC while the 8-ore CMP hs n 8MB LLC. The results show tht QBS improves performne of 4-ore CMP with n inlusive he by 8.% on verge. Non-inlusive nd exlusive he hierrhies for the 4-ore CMP improve performne by 8.3% nd.4% respetively. Similrly, for n 8-ore CMP, QBS improves verge performne by 9.5% while non-inlusive nd exlusive LLCs improve performne by 0.% nd 3.6% respetively. These results show tht QBS sles with inresing number of ores nd onsistently bridges more thn 95% of the performne gp between inlusive nd non-inlusive he hierrhies. H. Hrdwre Overhed The proposed TLA poliies require very little hrdwre overhed. They ll use existing dt pths in the system to llow the LLC to beome wre of the temporl lolity of lines in the ore hes. TLH requires new messge lss tht does not return dt but merely updtes the LLC replement stte. Sine TLH is not slble due to lrge bndwidth requirements, TLHs re not prtil solution espeilly sine ECI nd QBS serve s lterntive low bndwidth solutions. ECI requires dditionl logi to preselet the next vitim in the he nd invlidte it from the ore hes. Similrly, QBS uses the existing dt pth to determine whether or not he line n be evited from the LLC. The dditionl logi required is stte-mhine to send queries to the ore hes nd responses to the LLC. As the number of HW-threds on-hip ontinue to inrese, both QBS nd ECI re slble solutions tht re effetive for ny inlusive he hierrhy orgniztion. VI. RELATED WORK Mnging multi-level he hierrhies hs been n importnt reserh re for severl dedes [26, 24,, 7, 8]. Ber et l first introdued the onept of multi-level inlusive he hierrhies for simplifying he oherene [6]. Jouppi et l proposed exlusive hes to redue LLC onflit misses nd to lso inrese the effetive he pity of the hierrhy by not repliting lines in the LLC. Exlusive hes perform best when the sum of the ore hes is not signifintly lrger thn the size of the LLC [7, ]. Non-inlusive he hierrhies hve been proposed in the ontext of ggressive

11 prefething [0, 28] nd in the ontext of better he mngement poliies [28, 2]. Zhrn [27, 28] proposed the use of globl replement to improve the performne of inlusion. The proposed replement poliies were only evluted for single-threded worklods running in isoltion. The results showed tht while globl replement redued the number of inlusion vitims, the observed performne benefits were negligible. Our studies replited their results for single-threded worklods, however, our results indite tht globl temporl lolity indeed benefits CMPs. Grde et l [3] followed up on Zhrn s work nd deonstruted globl replement for single-ore nd multi-ore proessors. They evluted the potentil for globl replement in inlusive hes by nlyzing the miss strem of non-inlusive LLC. Speifilly, for their non-inlusive LLC, they mesured the number of times n LLC vitim ws resident in the ore hes upon evition (i.e. potentil inlusion vitim). The problem with there pproh is tht the potentil inlusion vitim ours only one for hot line beuse suh hot line ontinues to reeive hits in the ore hes nd is never re-filled in the LLC. The potentil for globl replement should hve been mesured in the presene of n inlusive he hierrhy. Flether et l [2] observed inlusion vitims in the ontext of diret mpped network hes nd proposed three solutions to ddress the problem. The proposed solutions inlude inresing the he ssoitivity, vitim he [8], or mking the LLC non-inlusive nd using snoop filter (lled tg he in the pper) to ese he oherene. They showed tht inresing the network he ssoitivity nd vitim hes redue the negtive effets of inlusion vitims. The proposed mehnisms require dditionl hrdwre strutures (e.g. vitim he nd snoop filter). Our work differs from the work of Flether et l in tht we do not require ny dditionl hrdwre strutures. The Erly Core Invlidtion (ECI) proposl effetively uses n in-llc vitim he insted of n externl vitim he. For our bseline system, we ompred the performne of both ECI nd QBS to n inlusive LLC bked by 32-entry vitim he. We found tht the 32-entry vitim he improves verge performne by only 0.8% while ECI nd QBS improves verge performne by 4.5% nd 6.5% respetively. There hs lso been extensive reserh on mnging shred hes in CMPs [4, 5, 25]. Most of the prior reserh work fouses on how to effiiently prtition the shred lstlevel he of CMP. When multiple pplitions ompete for the shred LLC, the proposed poliies dynmilly provide more he to pplitions tht benefit from the LLC nd less to pplitions tht do not. However, the proposed poliies only trget he prtitioning nd do not ddress the problem of inlusion vitims. We ompred the TLA poliies in the presene of intelligent he mngement poliies [4, 5] nd find tht we hieve similr performne improvements. VII. SUMMARY Inlusive hes re desirble beuse they simplify he oherene. However, inlusive hes limit performne due to inlusion vitims. In ontrst, non-inlusive hes eliminte inlusion vitims but ome t the expense of inresing oherene omplexity. This pper improves inlusive he performne by mking these ontributions:. We show tht better mnged inlusive he provides equl or better performne thn non-inlusive he. Speifilly, we show tht the bulk of non-inlusive (nd exlusive) he performne over inlusive hes is due to voiding inlusion vitims nd not the extr hing pity in the hierrhy. Inlusion vitims our beuse n inlusive LLC is unwre of the temporl lolity of hot lines in the ore hes. 2. We propose Temporl Lolity Hints (TLH) s mehnism to onvey the temporl lolity of lines referened in the ore hes to the LLC. We show tht TLHs sent by the L he signifintly redue the number of inlusion vitims. TLHs, however, signifintly inrese on-hip bndwidth. 3. We propose Erly Core Invlidtion (ECI) s low bndwidth tehnique used by the LLC to derive line s temporl lolity. Unlike TLHs, ECI is only triggered on LLC misses. ECI selets the next potentil LLC vitim nd invlidtes tht line from the ore hes while retining it in the LLC. ECI employs the heuristi tht if the next potentil LLC vitim line is hot, it will be servied by the LLC (insted of the ore hes) on the subsequent ess using n LLC replement stte updte. However, ECI limits performne when the subsequent ess ours fter miss (to the sme set). 4. We propose Query Bsed Seletion (QBS) s n lterntive to ECI. QBS employs the heuristi tht the ore hes re best suited to inform the LLC on the temporl lolity of line. In the pper, we show tht lines resident in the ore hes hve high temporl lolity nd should not be evited from the LLC. In doing so, QBS llows inlusive hes to pproh the performne of non-inlusive hes. For our bseline 2-ore CMP, we show tht TLHs issued from the L he bridge 85% of the gp between inlusive nd non-inlusive hes. ECI bridges 55% of the gp nd QBS performs similr to non-inlusive hes. Our studies show tht when non-inlusion must be enfored due to trget ore he to LLC size rtios, (e.g. :2 or :4), the QBS tehnique llows hip designers to mintin the oherene benefits of inlusion while performing similr to non-inlusive he. This is espeilly importnt sine hip designers tody re willing to srifie oherene benefits for the performne improvement of non-inlusive LLCs []. Using QBS, we report 0-33% performne improvement for 25 of the 05 worklods on 2, 4 nd 8 ore systems with :4 hierrhy nd 6% performne improvement on systems with :2 hierrhy. As the number of ores shring n LLC inreses, he ontention inreses nd ddressing inlusion vitims beomes growing onern. ECI nd QBS re slble nd perform better with inresing ore ounts. Both introdue extr messges into the system tht do not require signifint dditionl bndwidth or dditionl hrdwre strutures.

Total Rewards: Vacation, Sick and Personal Leave for Full Time Employees

Total Rewards: Vacation, Sick and Personal Leave for Full Time Employees POLICY: 6Hx28:3D-03 Responsible Exeutive: Vie President, Orgniztionl Development & Humn Resoures Poliy Contts: Diretor, HR Poliy nd Compline Progrms Speifi Authority: 1001.61, F.S. Lw Implemented: 1001.64,

More information

A Fast Heuristic Scheduling Algorithm for Periodic ConcurrenC Models

A Fast Heuristic Scheduling Algorithm for Periodic ConcurrenC Models A Fst Heuristi Sheduling Algorithm for Periodi ConurrenC Models Weiwei Chen nd Riner Doemer Center for Emedded Computer Systems University of Cliforni, Irvine Outline Introdution nd Relted Work ConurrenC

More information

Nonlinear Mixed Effects Model for Swine Growth

Nonlinear Mixed Effects Model for Swine Growth Nonliner Mixed Effects Model for Swine Growth A. P. Schinckel nd B. A. Crig Deprtment of Animl Sciences nd Deprtment of Sttistics, Purdue University Introduction Severl nonliner growth functions model

More information

6.1 Damage Tolerance Analysis Procedure

6.1 Damage Tolerance Analysis Procedure 6. Dmge Tolernce Anlysis Procedure For intct structure the nlysis procedures for Slow Crck Growth nd Fil Sfe structure re essentilly the sme. An initil flw is ssumed nd its growth is nlyzed until filure

More information

On Using Lossless Compression of Debug Data in Embedded Logic Analysis

On Using Lossless Compression of Debug Data in Embedded Logic Analysis On Using Lossless Compression of Deug Dt in Emedded Logi Anlysis Eh Anis nd Niol Niolii Deprtment of Eletril nd Computer Engineering MMster University, Hmilton, ON L8S 4K1, Cnd Emil: nise@mmster, niol@eemmster

More information

Three-Phase Wound-Rotor Induction Machine with a Short- Circuited Rotor

Three-Phase Wound-Rotor Induction Machine with a Short- Circuited Rotor Exercise 1 Three-Phse Wound-Rotor Induction Mchine with Short- Circuited Rotor EXERCISE OBJECTIVE When you hve completed this exercise, you will know how three-phse woundrotor induction mchine cn operte

More information

Crop Rotations, Reduced Tillage and N Fertilization Effects on Corn Yields And Aflatoxin Levels

Crop Rotations, Reduced Tillage and N Fertilization Effects on Corn Yields And Aflatoxin Levels Crop Rottions, Redued Tillge nd N Fertiliztion Effets on Corn Yields And Afltoxin Levels J.E. Mtoh nd M. Rihrdson Texs AgriLife Reserh nd Extension Center Stte Hwy Corpus Christi, TX 78- jmtoh@g.tmu.edu

More information

Table 2(a) Description of quantified economy-wide emission reduction target: base year a

Table 2(a) Description of quantified economy-wide emission reduction target: base year a Tle 2() Desription of quntified eonomy-wide emission redution trget: se yer Prty Bse yer /se period Emission redution trget Period for rehing trget % of se yer/se period % of 1990 Reporting y developed

More information

Implications of Accelerated Agricultural Growth on Household Incomes and Poverty in Ethiopia: A General Equilibrium Analysis

Implications of Accelerated Agricultural Growth on Household Incomes and Poverty in Ethiopia: A General Equilibrium Analysis ESSP2 Disussion Pper 002 Implitions of Aelerted Agriulturl Growth on Household Inomes nd Poverty in Ethiopi: A Generl Equilibrium Anlysis Pul Dorosh nd Jmes Thurlow with the support of the EDRI/University

More information

Sustainable Crop Rotations for Alberta s Brown Soil Zone

Sustainable Crop Rotations for Alberta s Brown Soil Zone Otober 2003 Agdex 515-1 Sustinble Crop Rottions for Albert s Brown Soil Zone C rop prodution on the Cndin Priries hs historilly foused on spring whet. In reent dedes, rops suh s brley nd nol hve inresed

More information

Small Business Cloud Services

Small Business Cloud Services Smll Business Cloud Services Summry. We re thick in the midst of historic se-chnge in computing. Like the emergence of personl computers, grphicl user interfces, nd mobile devices, the cloud is lredy profoundly

More information

2016 Prelim Essay Question 2

2016 Prelim Essay Question 2 216 Prelim Essy Question 2 In recent yers, the price of nturl fertilisers for orgnic brown rice production hs risen nd helthy living cmpigns re seeing more consumers switching from nonorgnic white rice

More information

Chapter 9. Quadratics

Chapter 9. Quadratics Chpter 9 Qudrtics Artificil Body Prts 9.1 Solving Qudrtic Equtions by Fctoring 9. Completing the Squre 9.3 The Qudrtic Formul 9.4 Eponentil Functions (Growth nd Decy) Chpter Review Chpter Test 147 Section

More information

RAIN (RAndom INsertion) Scheduling Algorithm for SoC Test

RAIN (RAndom INsertion) Scheduling Algorithm for SoC Test RAIN (RAnom INsertion) Sheuling Algorithm for SoC Test Jung-Been Im Sunghoon Chun Geune Kim Jin-Ho An Sungho Kng Deprtment of Eletril n Eletroni Engineering Yonsei University 134, Shinhon-Dong Seoemoon-Gu,

More information

1 Information, Persuasion, and Signalling

1 Information, Persuasion, and Signalling ECON 312: Advertising 1 We will now exmine nother strtegic vrible vilble to firms, tht of dvertising. Industril Orgniztion Advertising 1 Informtion, Persusion, nd Signlling 1.1 Persusion versus Informtion

More information

SHEAR STRENGTH OF CONCRETE-ENCASED COMPOSITE STRUCTURAL MEMBERS

SHEAR STRENGTH OF CONCRETE-ENCASED COMPOSITE STRUCTURAL MEMBERS SHEAR STRENGTH OF CONCRETE-ENCASED COMPOSITE STRUCTURAL MEMBERS By C. C. Weng, 1 S. I. Yen, 2 nd C. C. Chen 3 ABSTRACT: This study investigtes the sher strength of omposite steel nd onrete members in whih

More information

Effect of Nitrogen Rate on Yield of Nine Warm-season Introduced Perennial Forage Varieties

Effect of Nitrogen Rate on Yield of Nine Warm-season Introduced Perennial Forage Varieties Effet of Nitrogen Rte on Yield of Nine Wrm-seson Introdued Perennil Forge Vrieties By Eddie Funderurg, Jon T. Biermher, Corey Moffet, Mohu Hque nd Jgdeesh Mosli When rnhers think out plnting n introdued

More information

EVALUATION OF ALTERNATIVE FUNGICIDES FOR CONTROL OF CERCOSPORA SPOT ON FUERTE

EVALUATION OF ALTERNATIVE FUNGICIDES FOR CONTROL OF CERCOSPORA SPOT ON FUERTE Proeedings V World Avodo Congress (Ats V Congreso Mundil del Agute) 23. pp. 579-583. EVALUATION OF ALTERNATIVE FUNGICIDES FOR CONTROL OF CERCOSPORA SPOT ON FUERTE A Willis nd JA Duvenhge Merensky Tehnologil

More information

Chapter 02 - Putting the Customer First

Chapter 02 - Putting the Customer First 1. About hlf of every dollr tht consumers spend on products pys for mrketing costs. LEARNING OBJECTIVES: SEM.KO.4.LO: 2.1-1 - LO: 2.1-1 2. The mrketing concept requires mintennce of importnt reltionships

More information

Physiological effects of mechanized harvesting and ways to minimize its impacts

Physiological effects of mechanized harvesting and ways to minimize its impacts 8/3/218 Physiologil effets of mehnized hrvesting nd wys to minimize its impts S.R.W. Pthirnge 1, M.A. Wijertne 1 nd W.A.J.M. De Cost 2 1 Agronomy Division, Te Reserh Institute of Sri Lnk 2 Fulty of Agriulture,

More information

Modelling and Prediction of NOx emissions from coal fired boilers: Case study

Modelling and Prediction of NOx emissions from coal fired boilers: Case study Interntionl Reserh Journl of Engineering nd Tehnology (IRJET) e-issn: 9 - Volume: Issue: June- www.irjet.net p-issn: 9-7 Modelling nd redition of NO emissions from ol fired boilers: se study Vineeth Morris

More information

Preparation of conductive carbons with high surface area

Preparation of conductive carbons with high surface area PERGAMON Crbon 39 (2001) 39 44 Preprtion of ondutive rbons with high surfe re Weiming Lu, D.D.L. Chung* Composite Mterils Reserh Lbortory, Stte University of New York t Bufflo, Bufflo, NY 14260-4400, USA

More information

Practices and Strategic Investment Section

Practices and Strategic Investment Section FOREST PRACTICES Prties nd Strtegi Investment Setion Resoure Prties Brnh PO Box 9513 Stn Prov. Govt. Vitori. BC V8W 9C2 SILVICULTURE NOTE 3 July 211 Twenty-Yer Effets of Windrow Burning, Chemil nd Mehnil

More information

[ HOCl] Chapter 16. Problem. Equilibria in Solutions of Weak Acids. Equilibria in Solutions of Weak Acids

[ HOCl] Chapter 16. Problem. Equilibria in Solutions of Weak Acids. Equilibria in Solutions of Weak Acids Equilibri in Solutions of Wek Acids Chpter 16 Acid-Bse Equilibri Dr. Peter Wrburton peterw@mun.c http://www.chem.mun.c/zcourses/1011.php The dissocition of wek cid is n equilibrium sitution with n equilibrium

More information

Three-Phase Wound-Rotor Induction Machine with Rotor Resistance

Three-Phase Wound-Rotor Induction Machine with Rotor Resistance Exercise 2 Three-Phse Wound-Rotor Induction Mchine with Rotor Resistnce EXERCISE OBJECTIVE When you hve completed this exercise, you will know the effects of vrying the rotor resistnce of three-phse wound-rotor

More information

p Coaches j i n C Dimensions Report Name Ali Example Date of Report: 29/06/2016 Team Profile 3

p Coaches j i n C Dimensions Report Name Ali Example Date of Report: 29/06/2016 Team Profile 3 Report Nme Ali Exmple Dte of Report: 29/06/2016 Tem Profile 3 Also Recommended: Behviourl Type t Work Profile, Composite Tem Report Who could use components of this report: p Coches j i HR professionls

More information

Job Description. Senior Lecturer. Electronic & Electrical Engineering. Education and Research. Head of Department/Group. Any research staff/students

Job Description. Senior Lecturer. Electronic & Electrical Engineering. Education and Research. Head of Department/Group. Any research staff/students Jo Desription Jo title Deprtment/Shool Jo fmily Senior Leturer Eletroni & Eletril Engineering Edution nd Reserh Grde 9 Reporting to Responsile for Lotion Hed of Deprtment/Group Any reserh stff/students

More information

EMPLOYER HUB APPRENTICESHIPS THE. 2019/20 Apprenticeships. Information for applicants. The college of choice

EMPLOYER HUB APPRENTICESHIPS THE. 2019/20 Apprenticeships. Information for applicants. The college of choice EMPLOYER HUB APPRENTICESHIPS THE 2019/20 Apprenticeships Informtion for pplicnts The college of choice Contents Why Nescot? 04 Wht is n pprenticeship? 04 Wht re the benefits of doing n pprenticeship?

More information

PROCEEDINGS 2017 Crop Pest Management Short Course & Minnesota Crop Production Retailers Association Trade Show

PROCEEDINGS 2017 Crop Pest Management Short Course & Minnesota Crop Production Retailers Association Trade Show PROCEEDINGS 2017 Crop Pest Mngement Short Course & Minnesot Crop Prodution Retilers Assoition Trde Show Institute for Ag Professionls http://www.extension.umn.edu/griulture/g-professionls/ Do not reprodue

More information

Belief-Revisions After Earnings Announcements: Evidence From Security Analysts Forecast Revisions

Belief-Revisions After Earnings Announcements: Evidence From Security Analysts Forecast Revisions Belief-Revisions After Ernings Announements: Evidene From Seurity Anlysts Forest Revisions Ping Eri Yeung Deprtment of Aounting Chrles H. Lundquist College of Business University of Oregon Eugene, OR 9743-18

More information

Business Continuity Software Buyer s Guide

Business Continuity Software Buyer s Guide Business Continuity Softwre Buyer s Guide Opertionlly strtegic nd deployble, business continuity plns re criticl to ensuring your orgniztion cn survive nd succeed following n unplnned incident. Mny orgniztions

More information

NOTICE CONCERNING COPYRIGHT RESTRICTIONS

NOTICE CONCERNING COPYRIGHT RESTRICTIONS NOTICE CONCERNING COPYRIGHT RESTRICTIONS This document my contin copyrighted mterils. These mterils hve been mde vilble for use in reserch, teching, nd privte study, but my not be used for ny commercil

More information

CLOUD-EXTENSIBLE TRANSCODING AVOID THE TRAP OF SINGLE CLOUD OVER-DEPENDENCY

CLOUD-EXTENSIBLE TRANSCODING AVOID THE TRAP OF SINGLE CLOUD OVER-DEPENDENCY CLOUD-EXTENSIBLE TRANSCODING AVOID THE TRAP OF SINGLE CLOUD OVER-DEPENDENCY compny Cloud-Extensible Trnscoding TABLE OF CONTENTS Trnscode Worklod Chllenges... 3 Cloud Dependency... 4 Cloud-Extensibility...

More information

The Effect of SFAS No. 131 on the Diversification Discount

The Effect of SFAS No. 131 on the Diversification Discount The Effect of SFAS No. 131 on the Diversifiction Discount Seoungpil Ahn Sogng Business School, Sogng University PA706, 35 Bekbeom-ro, Mpo-gu, Seoul 121-742, Kore E-mil: sphn@sogng.c.kr Received: July 2,

More information

(b) Is already deposited in a waste disposal site without methane recovery.

(b) Is already deposited in a waste disposal site without methane recovery. TYPE III - OTHER PROJECT ACTIVITIES Project prticipnts must tke into ccount the generl guidnce to the methodologies, informtion on dditionlity, bbrevitions nd generl guidnce on lekge provided t http://cdm.unfccc.int/methodologies/sscmethodologies/pproved.html.

More information

sensitive VBSs Vh subdomains EF EF

sensitive VBSs Vh subdomains EF EF Tlin- Tlin-EGFP-His 2 3 2 3 ABD Mehno- ABD2 ABD3 sensitive VBSs DD 6xHis EGFP Vh sudomins EGFP-Vinulin EGFP 2 3 4 Vt EGFP-Vh EGFP 2 3 4 -tinin- CH CH2 SP SP SP SP EF EF mcherry--tinin- mcherry CH CH2 SP

More information

Broken rice facts. Physical and Functional Characteristics of Broken Rice Kernels. Parboiling Process. May 23, Production of broken rice 14-10%

Broken rice facts. Physical and Functional Characteristics of Broken Rice Kernels. Parboiling Process. May 23, Production of broken rice 14-10% My 23, 218 Physil nd Funtionl Chrteristis of Broken Rie Kernels Broken rie fts Redued eonomi vlue Undesirle Inevitle Ree M. Brue University of Arknss Advisor: Dr. Griffiths G. Atungulu Inexpensive Affets

More information

MR405: Response of Young Black Spruce (Picea mariana (Mill.) B.S.P.) to a Mixture of Wood Ash and Secondary Papermill Sludge

MR405: Response of Young Black Spruce (Picea mariana (Mill.) B.S.P.) to a Mixture of Wood Ash and Secondary Papermill Sludge The University of Mine DigitlCommons@UMine Misellneous Reports Mine Agriulturl nd Forest Experiment Sttion -997 MR45: Response of Young Blk Sprue (Pie mrin (Mill.) B.S.P.) to Mixture of Wood Ash nd Seondry

More information

Installation Instructions for Trantorque Keyless Bushings

Installation Instructions for Trantorque Keyless Bushings Instlltion Instrutions for Trntorque Keyless Bushings A Trntorque Keyless Bushing offers flexile n esy instlltion while proviing exeptionl holing power. Referring to the series, plese follow these Instlltion

More information

CHAPTER 2 RELATIONAL MODEL

CHAPTER 2 RELATIONAL MODEL CHAPTER RELATIONAL MODEL Chpter : Reltionl Model Structure of Reltionl Dtbses Fundmentl Reltionl Algebr Opertions Additionl Reltionl Algebr Opertions Extended Reltionl-Algebr-Opertions Null Vlues Modifiction

More information

Linked List Implementation of Discount Pricing in Cloud

Linked List Implementation of Discount Pricing in Cloud Linked List Implementtion of Discount Pricing in Cloud Mlr.J 1, S.Jesinth Strvin 2 Mster of Engineering in computer Science, Assistnt professor IT deprtment Ponjesly College of Engineering Abstrct: In

More information

Chrome River Job Aid Reconciling PCard in Chrome River

Chrome River Job Aid Reconciling PCard in Chrome River Chrome River Jo Ai Reoniling PCr in Chrome River 1. Log into Chrome River. 2. Open e-wllet to etermine types of trnstions to e reonile. From the Home sreen:. Selet the hmurger menu ion. Selet ewllet. The

More information

Katherine J. Hohnholt

Katherine J. Hohnholt Coneptul Design of uler Systems for Hydrogen Prodution by Ktherine J. Hohnholt Submitted to the Deprtment of uler Siene nd Engineering in prtil fulfillment of the requirements for the degree of Bhelor

More information

p Coaches j i m Recruitment Dimensions Report Name Ali Example Date of Report: 29/06/2016 Elements report 3

p Coaches j i m Recruitment Dimensions Report Name Ali Example Date of Report: 29/06/2016 Elements report 3 Report Nme Ali Exmple Dte of Report: 29/06/2016 Elements report 3 Also Recommended: Trit Profile, Competency Report Who could use components of this report: p Coches j i HR professionls Trined prctitioners

More information

Quantifying the Total Cost of Ownership for Entry-Level and Mid-Range Server Clusters

Quantifying the Total Cost of Ownership for Entry-Level and Mid-Range Server Clusters Quntifying the Totl Cost of Ownership for Entry-Level nd Mid-Rnge Server Clusters A Detiled Anlysis of the Totl Cost of Ownership of OpenVMS, IBM AIX nd Sun Solris server clusters. June 2007 Version 1.0

More information

Model 3500 Transmitter (MVD) or Model 3300 Peripheral

Model 3500 Transmitter (MVD) or Model 3300 Peripheral Quik Referene Guide P/N 20000993, Rev. B August 2005 Model 3500 Trnsmitter (MVD) or Model 3300 Peripherl Instlltion Instrutions for Rk-Mount For tehnil support, phone the support enter nerest you: In the

More information

Introduction. and Hispanics as $1.3 trillion.

Introduction. and Hispanics as $1.3 trillion. Introduction Supplier diversity is the proctive business progrm supporting minority owned, women owned, vetern owned, LGBT owned, service disbled vetern owned, historiclly underutilized business nd SBA

More information

Organic Cover Crop Research at WSU Puyallup

Organic Cover Crop Research at WSU Puyallup Orgnic Cover Crop Reserch t WSU Puyllup Ferury 8 Crig Cogger, Andy Bry, nd Liz Myhre Wshington Stte University Puyllup Reserch nd Extension Center Astrct: Cover crops re loclly grown source of orgnic mtter

More information

Copyright 1982 by ASME. Combined Cycles

Copyright 1982 by ASME. Combined Cycles THE AMERICAN OCIETY OF MECHANICAL ENGINEER 345 E. 47 t., New York, N.Y. 117 82-GT-38 ^,+ w The ociety shll not be responsible for sttements or opinions dvnced in ppers or in C discussion t meetings of

More information

Abstract # Strategic Inventories in a two-period Cournot Duopoly. Vijayendra Viswanathan Jaejin Jang. University of Wisconsin-Milwaukee

Abstract # Strategic Inventories in a two-period Cournot Duopoly. Vijayendra Viswanathan Jaejin Jang. University of Wisconsin-Milwaukee Abstrct # 0-0788 Strtegic Inventories in two-period Cournot Duopoly Vijyendr Viswnthn Jejin Jng University of Wisconsin-Milwukee P.O. Box 784 Deprtment of Industril nd Mnufcturing Engineering University

More information

Technology & Prod uct Reports Does Amendment of Soak Solution with Sucrose and Urea Increase Production of Shiitake Mushrooms on Sawdust Blocks?

Technology & Prod uct Reports Does Amendment of Soak Solution with Sucrose and Urea Increase Production of Shiitake Mushrooms on Sawdust Blocks? Tehnology & Prod ut Reports Does Amendment of Sok Solution with Surose nd Ure Inrese Prodution of Shiitke Mushrooms on Swdust Bloks? Cthy Sot, Cul Beyl, nd Gokul Ghle ADDITIONAL INDEX WORDS. iologil effiieny,

More information

Primer in Population Genetics

Primer in Population Genetics Primer in Popultion Genetics Hierrchicl Orgniztion of Genetics Diversity Primer in Popultion Genetics Defining Genetic Diversity within Popultions Polymorphism number of loci with > 1 llele Number of lleles

More information

COMPARISON OF NANOCRYSTALLINE FRACTIONS DETERMINED BY VARIOUS METHODS

COMPARISON OF NANOCRYSTALLINE FRACTIONS DETERMINED BY VARIOUS METHODS 476 Rev.Adv.Mter.Si. 18(2008) 476-480 K. Pêk³, P. Jskiewiz, A. Grbis nd J. Ltuh COMPARISON OF NANOCRYSTALLINE FRACTIONS DETERMINED BY VARIOUS METHODS K. Pêk³ 1, P. Jskiewiz 1, A. Grbis 2 nd J. Ltuh 3 1

More information

Comparison of two soil quality indexes to evaluate cropping systems in northern Colorado

Comparison of two soil quality indexes to evaluate cropping systems in northern Colorado University of Nebrsk - Linoln Digitlommons@University of Nebrsk - Linoln Publitions from USD-RS / UNL Fulty USD griulturl Reserh Servie --Linoln, Nebrsk 1-1-28 omprison of two soil qulity indexes to evlute

More information

Report to the Southwest Florida Water Management District. Effects of Microsprinkler Irrigation Coverage on Citrus Performance

Report to the Southwest Florida Water Management District. Effects of Microsprinkler Irrigation Coverage on Citrus Performance Report to the Southwest Florid Wter Mngement District Effects of Microsprinkler Irrigtion Coverge on Citrus Performnce L. R. Prsons University of Florid Institute of Food nd Agriculturl Sciences Citrus

More information

Effects of herbaceous vegetation control and aspen stem density on boreal mixedwood stand development. Partners Report Field Season

Effects of herbaceous vegetation control and aspen stem density on boreal mixedwood stand development. Partners Report Field Season www.forestreserh., Produt tlogue, Reports, RP-38 Effets of hereous vegettion ontrol nd spen stem density on orel mixedwood stnd development. Prtners Report - 26 Field Seson (FRP projet 13-31 (6331); SERG-I

More information

PRODUCTION METHOD. contact sulphuric acid plant, various kind of waste gases).

PRODUCTION METHOD. contact sulphuric acid plant, various kind of waste gases). "POL" PHOSPHATE FERTLZER DRY AND WASTE-FREE PRODUCTON METHOD / The elborted fertilizer prodution tehnology nd POL -phosphte fertilizer, re originl hievements dpted to present hnges in the rw-mteril nd

More information

Platform quick guide. Stockbroking Pro platform. stockbroking

Platform quick guide. Stockbroking Pro platform. stockbroking Stockbroking Pro pltform Pltform quick guide This concise guide hs been put together to help you quickly fmilirise yourself with the mny fetures nd tools vilble on the CMC Mrkets Stockbroking Pro pltform.

More information

The method of setting of heat consumption standard for buildings

The method of setting of heat consumption standard for buildings The method o setting o het onsumption stndrd or buildings Levtsev Alexey Pvlovih Dotor o Tehnil Sienes Proessor Hed o Deprtment «Het nd Power Systems» Ntionl Reserh Mordovi Stte University o Srns 68 Bolshevistsy

More information

We re like money in the bank.

We re like money in the bank. With our experiene, knowledge, and expertise, we re more than just a great way to measure effetiveness, maintain exellene, and identify opportunities for improvement. We re like money in the bank. Ahieving

More information

European Treaty Series - No. 158 ADDITIONAL PROTOCOL TO THE EUROPEAN SOCIAL CHARTER PROVIDING FOR A SYSTEM OF COLLECTIVE COMPLAINTS

European Treaty Series - No. 158 ADDITIONAL PROTOCOL TO THE EUROPEAN SOCIAL CHARTER PROVIDING FOR A SYSTEM OF COLLECTIVE COMPLAINTS Europen Trety Series - No. 158 ADDITIONAL PROTOCOL TO THE EUROPEAN SOCIAL CHARTER PROVIDING FOR A SYSTEM OF COLLECTIVE COMPLAINTS Strsbourg, 9.XI.1995 2 ETS 158 - Europen Socil Chrter (Additionl Protocol),

More information

Omega 41 (2013) Contents lists available at SciVerse ScienceDirect. Omega. journal homepage:

Omega 41 (2013) Contents lists available at SciVerse ScienceDirect. Omega. journal homepage: Omeg 41 (13) 75 765 Contents lists vilble t SiVerse SieneDiret Omeg journl homepge www.elsevier.om/lote/omeg Fresh-produt supply hin mngement with logistis outsouring Xioqing Ci Jin Chen bn Yongbo Xio

More information

Product design. A product is a bundle of attribute levels or features that have utilities to customer (price is considered as attribute as well)

Product design. A product is a bundle of attribute levels or features that have utilities to customer (price is considered as attribute as well) Product design Working ssumption: Wht is product? A product is undle of ttriute levels or fetures tht hve utilities to customer (price is considered s ttriute s well) The mening of : Designing product

More information

Employed Worker Training Agreement

Employed Worker Training Agreement Employe Worker Trining Agreement SECTION : GENERAL INFORMATION Orgniztion Nme: Street Aress: Authorize Contt Person: Telephone Numer: Emil Aress: Zip Coe: Fx Numer: Wesite Aress: Dte of Estlishment: Yers

More information

Pre-Plant Broadcast Urea in Direct Seeding, A Logistical Return to the Past? Tom Jensen

Pre-Plant Broadcast Urea in Direct Seeding, A Logistical Return to the Past? Tom Jensen Pre-Plnt Brodcst Ure in Direct Seeding, A Logisticl Return to the Pst? Tom Jensen Interntionl Plnt Nutrition Institute (IPNI), Northern Gret Plins Director 102-411 Downey Rd., Ssktoon, SK S7N 4L8 Ph: 306-652-3467

More information

The basic model for inventory analysis

The basic model for inventory analysis The bsic model for inventory nlysis Lecture Notes for ME515 Prepred by Joyce Smith Cooper Professor of Mechnicl Engineering University of Wshington cooperjs@uw.edu See Chpter 2 of Heijungs nd Suh (22)

More information

H. Randall Smith; Ph.D. Agronomy and Wayne Porter: Ph.D. Horticulture Mississippi State University Extension Service

H. Randall Smith; Ph.D. Agronomy and Wayne Porter: Ph.D. Horticulture Mississippi State University Extension Service Effect of SumGrow on growth, development nd yield of Irish pottoes (Solnum tuerosum) t the Beumont Reserch Sttion (Mississippi Stte University) during 217 H. Rndll Smith; Ph.D. Agronomy nd yne Porter:

More information

Productivity convergence in OECD manufacturing industries

Productivity convergence in OECD manufacturing industries Eonomis Letters 66 (000) 337 345 www.elsevier.om/ lote/ eonse Produtivity onvergene in OECD mnufturing industries,,,,d M.A. Crree *, L. Klomp, A.R. Thurik Fulty of Eonomis, Centre for Advned Smll Business

More information

7 mm Diameter Miniature Single-Turn Cermet Trimmer

7 mm Diameter Miniature Single-Turn Cermet Trimmer 7 mm Dimeter Miniture Single-Turn Cermet Trimmer A dust seled plsti se proteting qulity ermet trk gurntees high performne nd proven reliility. Adjustments re mde esier y the ler sle redings. is idelly

More information

Life cycle management -Prediction -

Life cycle management -Prediction - JST Oen Seminr Soro, 2 Mrh 2 Life yle mngement -Predition - Professor, Hokkido University, Jn Hiroshi YOKOTA, PhD, PE Life-Cyle Mngement system Servie eriod Usge Design Environment Servie life Future ln

More information

SOIL-WATER CHARACTERISTIC CURVES AND DUAL POROSITY OF SAND DIATOMACEOUS EARTH MIXTURES

SOIL-WATER CHARACTERISTIC CURVES AND DUAL POROSITY OF SAND DIATOMACEOUS EARTH MIXTURES SOIL-WATER CHARACTERISTIC CURVES AND DUAL POROSITY OF SAND DIATOMACEOUS EARTH MIXTURES By Crig A. Burger 1 nd Chrles D. Shkelford, 2 Assoite Memer, ASCE ABSTRACT: The soil-wter hrteristi urves (SWCCs)

More information

Combined Hot Air-Microwave Drying Methods in Banana Chips Production

Combined Hot Air-Microwave Drying Methods in Banana Chips Production Amerin-Eursin J. Agri. & Environ. Si., (8): 77-776, ISSN 88-6769 IDOSI Pulitions, DOI:.589/idosi.ejes...8.88 Comined Hot Air-Mirowve Drying Methods in Bnn Chips Prodution Hmid Tvkolipour nd Leil Zirjni

More information

The Retail Ombudsman complaint form

The Retail Ombudsman complaint form The Retil Ombudsmn complint form Welcome to our retil complints form. To proceed with your complint plese follow the 6 steps below nd provide ll of the informtion reuested. To be eligible to mke complint

More information

Steels used in braking systems

Steels used in braking systems Steels use in rking systems A report A mterils engineer is require to prepre report on the seletion of plin ron steels for use in the proution of vrious omponents for rke mnufturing ompny. Portions of

More information

TTT DIAGRAM OF A NEWLY DEVELOPED NICKEL-BASE SUPERALLOY ALLVAC 718PLUS

TTT DIAGRAM OF A NEWLY DEVELOPED NICKEL-BASE SUPERALLOY ALLVAC 718PLUS Superlloys 718, 625, 706 nd Derivtives 2005 Edited y E.A. Lori TMS (The Minerls, Metls & Mterils Soiety), 2005 TTT DIAGRAM OF A NEWLY DEVELOPED NICKEL-BASE SUPERALLOY ALLVAC 718PLUS Xishn Xie 1, Chunmei

More information

Received: 9 December 2008 / Accepted: 14 July 2009 / Published online: 14 August 2009 Ó U.S. Government 2009

Received: 9 December 2008 / Accepted: 14 July 2009 / Published online: 14 August 2009 Ó U.S. Government 2009 New Forests (2010) 39:195 213 DOI 10.1007/s11056-009-9164-5 Effet of midstory nd understory removl on the estblishment nd development of nturl nd rtifiil pin ok dvne reprodution in bottomlnd forests Jonthn

More information

Thermal Impact of Operating Conditions on the Performance of a Combined Cycle Gas Turbine

Thermal Impact of Operating Conditions on the Performance of a Combined Cycle Gas Turbine herml Impt of Operting Conditions on the Performne of Combined Cyle Gs urbine hmir K Ibrhim*, MM Rhmn, Fulty of Mehnil Engineering Universiti Mlysi Phng 6600 Pekn, Phng, Mlysi *thmirmthd@yhooom Automotive

More information

Pig breeding, selection and hybridisation in Italy

Pig breeding, selection and hybridisation in Italy Pig breeding, selection nd hybridtion in Itly Russo V. in Aumître A. (ed.). The production pig met in Mediterrnen Countries Pr : CIHEAM Options Méditerrnéennes : Série Etudes; n. 1989-I 1989 pges 91-97

More information

PF-PT N MEMBRANE FILTER ELEMENTS FEATURES & BENEFITS. Compressed Air & Process Filtration PF-PT N

PF-PT N MEMBRANE FILTER ELEMENTS FEATURES & BENEFITS. Compressed Air & Process Filtration PF-PT N PF-PT N MEMBRANE FILTER ELEMENTS Compressed Air & Proess Filtrtion The Donldson LifeTe PF-PT N filter element is sterile grde, pleted high performne PTFE memrne filter. It provides the gretest ssurne of

More information

The point at which quantity demanded and quantity supplied come together is known as equilibrium. Price of a slice of pizza $2.00. Demand $2.50 $3.

The point at which quantity demanded and quantity supplied come together is known as equilibrium. Price of a slice of pizza $2.00. Demand $2.50 $3. Blncing the Mrket The point t which quntity demnded nd quntity supplied come together is known s equilibrium. Finding Equilibrium per slice $3.5 $3. $2.5 $2. $1.5 $1. $.5 Equilibrium Point Equilibrium

More information

SLASH PINE FAMILIES IDENTIFIED WITH HIGH RESISTANCE TO FUSIFORM RUST. C. H. Walkinshaw '

SLASH PINE FAMILIES IDENTIFIED WITH HIGH RESISTANCE TO FUSIFORM RUST. C. H. Walkinshaw ' SLASH PINE FAMILIES IDENTIFIED WITH HIGH RESISTANCE TO FUSIFORM RUST C. H. Wlkinshw ' Abstrct.--Fusiform rust redily kills slsh pine, Pinus elliottii Engelm. vr. elliottii. When the number of rust-infected

More information

Stainless Steel Floor Channels - Standard Channel Product Tables

Stainless Steel Floor Channels - Standard Channel Product Tables Stinless Steel Floor Chnnels - Stndrd Chnnel Produt Tles Hrmer Stinless Steel stndrd floor hnnels re used to drin wste wter from the floor nd dishrge it diretly in to the sewge system. The high qulity

More information

Modelling renewable energy policies. d Artis Kancs. 1 Introduction

Modelling renewable energy policies. d Artis Kancs. 1 Introduction Modelling renewble energy poliies d Artis Kns Msshusetts Institute of Tehnology MA-02139, Cmbridge E-mil: kns@mit.edu Abstrt: The preprtion, implementtion, oordintion nd verifition of poliy mesures is

More information

SpeedCoach OPERATION. Operating Manual THANK YOU... CONTENTS

SpeedCoach OPERATION. Operating Manual THANK YOU... CONTENTS Operting nul THA YOU... FOR PURCHASING the system. We hve designed the system to provide you with ll the si informtion you need to trin effetively speed, distne, elpsed time, stroke rte nd stroke ount,

More information

Optimising soil disturbance for erosion and runoff control in row crops.

Optimising soil disturbance for erosion and runoff control in row crops. Ref: C0672 Optimising soil disturne for erosion nd runoff ontrol in row rops. Jonn Niziolomski, Ro Simmons, Jne Rikson nd Mike Hnn, Crnfield Soil nd Agri- Food Institute, Crnfield University, Crnfield,

More information

recessive lozenge-shaped-fly-eye "alleles" in trans: recessive lozenge-shaped-fly-eye "alleles" in trans:

recessive lozenge-shaped-fly-eye alleles in trans: recessive lozenge-shaped-fly-eye alleles in trans: Wht do we men (wht hve we ment) y " gene": Reding for lectures 15-17 (We F27, Fr F29, We M5) Chp 8: from 258 (Nonoverlpping...) to 261 ( Crcking) & from 285 (8.6) to 293 (end of "essentil concepts) Chp

More information

THE INFLUENCE OF THERMOMECHANICAL TREATMENT AND CHEMICAL COMPOSITION ON RECRYSTALLIZATION OF Al-Mg ALLOYS

THE INFLUENCE OF THERMOMECHANICAL TREATMENT AND CHEMICAL COMPOSITION ON RECRYSTALLIZATION OF Al-Mg ALLOYS Assoition of Metllurgil Engineers of Seri Sientifi pper AMES UDC:669.715 721.065.53.040.2-174=20 THE INFLUENCE OF THERMOMECHANICAL TREATMENT AND CHEMICAL COMPOSITION ON RECRYSTALLIZATION OF Al-Mg ALLOYS

More information

EXPLORING BIOFUMIGATIONAL POTENTIAL OF MUSTARDS

EXPLORING BIOFUMIGATIONAL POTENTIAL OF MUSTARDS EXPLORING BIOFUMIGATIONAL POTENTIAL OF MUSTARDS Oleg Dugovish*, (University of Cliforni Coopertive Extension - Ventur County), Jmes Downer nd Ole Beker (University of Cliforni Coopertive Extension -Ventur

More information

Harmony/ESW An Agile Real-time Development Process. Jeff Vodov

Harmony/ESW An Agile Real-time Development Process. Jeff Vodov Hrmony/ESW An Agile Rel-time Development Process Jeff Vodov Objectives Provide high-level overview of Hrmony Chllenges of SE nd SW Development Hrmony Best Prctices Rtionl / Telelogic Process Rodmp Provide

More information

Chickpeas Respond Well To Inoculation With TagTeam

Chickpeas Respond Well To Inoculation With TagTeam Chickpes Respond Well To Inocultion With TgTem S.M. Phelps, nd E. Hgele Philom Bios Inc., 318-111 Reserch Drive, Ssktoon, SK S7N 3R2 Abstrct Rhizobi strins were tested in TgTem pet nd grnule formultions

More information

2014 Southeast Hay Convention

2014 Southeast Hay Convention 214 Southest Hy Convention Effet of Polymer-Cote Ure on Bermugrss Forge Proution Effet of Polymer-Cote Ure on Bermugrss Forge Proution Introution Without A, users of fe risky lterntives. - H 3 voltiliztion

More information

home water audit kit take a look Clackamas River Water Providers (CRWP) Saving Water Makes Cents! CRWP Public Education and Outreach Cooordinator

home water audit kit take a look Clackamas River Water Providers (CRWP) Saving Water Makes Cents! CRWP Public Education and Outreach Cooordinator tke look Clckms River Wter Providers (CRWP) Encourges you to tke look t your home wter system to identify wys you cn conserve wter nd be more efficient with our precious wter resource. FOR MORE INFORMATION:

More information

SEEDING CLOVERS OR GRASSES INTO OLDER ALFALFA BENEFITS AND HAZARDS ABSTRACT INTRODUCTION

SEEDING CLOVERS OR GRASSES INTO OLDER ALFALFA BENEFITS AND HAZARDS ABSTRACT INTRODUCTION SEEDING CLOVERS OR GRASSES INTO OLDER ALFALFA BENEFITS AND HAZARDS STANDS- Mick Cnevri1, Dn Putnm2, Brbr Reed3, Rchel Long4, Steve Orlo~, Tom Lnini6, nd Lrry Godfrey7 ABSTRACT Deciding wht to do with n

More information

P6.1. Magnetic position sensor with low coercivity material

P6.1. Magnetic position sensor with low coercivity material P6. Mgnetic position sensor with low coercivity mteril Jernce N., Frchon D. Moving Mgnet Technologies rue Christin Huygens, 25 Besnçon, Frnce. Introduction Mgnetic position sensors re widely used, especilly

More information

Assessing Impacts of Alternative Renewable Energy Strategies. d'artis Kancs and Hans Kremers. EERI Research Paper Series No 03/2002 ISSN:

Assessing Impacts of Alternative Renewable Energy Strategies. d'artis Kancs and Hans Kremers. EERI Research Paper Series No 03/2002 ISSN: EERI Eonomis nd Eonometris Reserh Institute Assessing Impts of Alterntive Renewble Energy Strtegies d'artis Kns nd Hns Kremers EERI Reserh Pper Series No 03/2002 ISSN: 2031-4892 EERI Eonomis nd Eonometris

More information

Effect of weed competition on growth characteristics of sunflower at different levels of nitrogen fertilizer

Effect of weed competition on growth characteristics of sunflower at different levels of nitrogen fertilizer Aville online t www.sholrsreserhlibrry.om Sholrs Reserh Librry Annls of Biologil Reserh, 212, 3 (11):5162-5168 (http://sholrsreserhlibrry.om/rhive.html) ISSN 976-1233 CODEN (USA): ABRNBW Effet of weed

More information

FINAL REPORT Preventing soil, water and nutrient limitations to sustainable, quality fruit production

FINAL REPORT Preventing soil, water and nutrient limitations to sustainable, quality fruit production FIAL REPORT Preventing soil, wter nd nutrient limittions to sustinle, qulity fruit prodution Prinipl Investigtors: Denise eilsen, Gerry eilsen, Eugene J. Hogue, Sunghee Guk, Agriulture nd Agri-Food Cnd,

More information

Effect of Transplant Size on Yields and Returns of Bell Peppers. Nathan Howard, Brent Rowell, and John C. Snyder Department of Horticulture

Effect of Transplant Size on Yields and Returns of Bell Peppers. Nathan Howard, Brent Rowell, and John C. Snyder Department of Horticulture Effect of Trnsplnt Size on Yields nd Returns of Bell Peppers Nthn Howrd, Brent Rowell, nd John C. Snyder Deprtment of Horticulture Introduction Bell peppers hve een mjor vegetle crop for frmers in western

More information

HUMAN RESOURCES MANAGEMENT REFORM TIME FRAME

HUMAN RESOURCES MANAGEMENT REFORM TIME FRAME Distribution: Restricted REPL.VII/4/R.10 1 October 2005 Originl: English Agend Item 10 English IFAD Consulttion on the Seventh Replenishment of IFAD s Resources Fourth Session Doh (Qtr), 1-2 October 2005

More information

SAMPLING INTENSITY AND NORMALIZATIONS: EXPLORING COST-DRIVING FACTORS IN NATIONWIDE MAPPING OF TREE CANOPY COVER

SAMPLING INTENSITY AND NORMALIZATIONS: EXPLORING COST-DRIVING FACTORS IN NATIONWIDE MAPPING OF TREE CANOPY COVER 21 Joint Meeting of the Forest Inventory nd Anlysis (FIA) Symposium nd the Southern Mensurtionists SAMPLIG ITESITY AD ORMALIZATIOS: EXPRIG COST-DRIVIG FACTORS I ATIOWIDE MAPPIG OF TREE CAOPY COVER John

More information