Utility-Based Partitioning :A Low-Overhead, High-Performance, Run-time Mechanism to Partition Shared Caches

Moinuddin K.Qureshi, Univ of Texas at Austin MICRO’2006

2007, 12, 05PAK, EUNJI

Outline

Introduction and Motivation Utility-Based Cache Partitioning Evaluation Scalable Partitioning Algorithm Related Work and Summary

Introduction

CMP and shared caches are common

Applications compete for the shared cache

Partitioning policies critical for high per-formance

Traditional policies: Equal (half-and-half) Performance isolation. No adaptation LRU Demand based. Demand ≠ benefit (e.g. streaming)

Background

Low Utility

High Utility

Saturating Utility

Utility Uab = Misses with a ways – Misses with b ways

Motivation

Improve performance by giving more cache to the application that benefits more from cache

Framework for UCP

Three components: Utility Monitors (UMON) per core Partitioning Algorithm (PA) Replacement support to enforce parti-

tions

I$

D$Core1

I$

D$Core2

SharedL2 cache

Main Memory

UMON1 UMON2PA

Utility Monitors (UMON)

For each core, simulate LRU policy using ATD Hit counters in ATD to count hits per recency

position LRU is a stack algorithm: hit counts utility

E.g. hits(2 ways) = H0+H1MTD

Set B

Set E

Set G

Set A

Set CSet D

Set F

Set H

ATD

Set B

Set E

Set G

Set A

Set CSet D

Set F

Set H

++++(MRU)H0 H1 H2…H15(LRU)

Dynamic Set Sampling (DSS) Extra tags incur hardware and power

overhead DSS reduces overhead [Qureshi,ISCA’06] 32 sets sufficient (analytical bounds) Storage < 2kB/UMON

MTD

ATD Set B

Set E

Set G

Set A

Set CSet D

Set F

Set H

++++(MRU)H0 H1 H2…H15(LRU)

Set B

Set E

Set G

Set A

Set CSet D

Set F

Set H

Set B

Set E

Set G

Set A

Set CSet D

Set F

Set H

Set BSet ESet G

UMON (DSS)

DSS Bounds with Analytical Model

Us = Sampled mean (Num ways allocated by DSS) Ug = Global mean (Num ways allocated by Global)

P = P(Us within 1 way of Ug)

By Cheb. inequality:P ≥ 1 – variance/n

n = number of sampled sets

In general, variance ≤ 3

Partitioning algorithm

Evaluate all possible partitions and select the best

With a ways to core1 and (16-a) ways to core2: Hitscore1 = (H0 + H1 + … + Ha-1) ---- from UMON1

Hitscore2 = (H0 + H1 + … + H16-a-1) ---- from UMON2

Select a that maximizes (Hitscore1 + Hitscore2)

Partitioning done once every 5 million cycles

Way Partitioning

Way partitioning support: [Suh+ HPCA’02, Iyer ICS’04] Each line has core-id bits On a miss, count ways_occupied in set by miss-

causing application

ways_occupied < ways_given

Yes No

Victim is the LRU line from other app

Victim is the LRU line from miss-caus-ing app

Evaluation Methodology

Configuration Two cores: 8-wide, 128-entry window Private L1s L2: Shared, unified, 1MB, 16-way LRU-based Memory: 400 cycles, 32 banks

Benchmarks Two-threaded workloads divided into 5 categories

Used 20 workloads (four from each type)

1.0 1.2 1.4 1.6 1.8 2.0

Weighted speedup for the baseline

Metrics

Weighted Speedup (default metric) perf = IPC1/SingleIPC1 + IPC2/SingleIPC2 correlates with reduction in execution time

Throughput perf = IPC1 + IPC2 can be unfair to low-IPC application

Hmean-fairness perf = hmean(IPC1/SingleIPC1, IPC2/SingleIPC2) balances fairness and performance

Results for weighted speedup

UCP improves average weighted speedup by 11%

Results for throughput

UCP improves average throughput by 17%

Results for hmean-fair-ness

UCP improves average hmean-fairness by 11%

Effect of Number of Sampled Sets

Dynamic Set Sampling (DSS) reduces overhead, not bene-

fits

8 sets16 sets32 setsAll sets

Scalability issues

Time complexity of partitioning low for two cores(number of possible partitions ≈ number of ways)

Possible partitions increase exponentially with cores

For a 32-way cache, possible partitions: 4 cores 6545 8 cores 15.4 million

Problem NP hard need scalable partitioning al-gorithm

Greedy Algorithm [Stone+ ToC ’92]

GA allocates 1 block to the app that has the max utility for one block. Repeat till all blocks allocated

Optimal partitioning when utility curves are convex

Pathological behav-ior for non-convex curves

Problem with Greedy Algo-rithm

0

10

20

30

40

50

60

70

80

90

100

0 1 2 3 4 5 6 7 8

A

B

In each iteration, the utility for 1 block:

U(A) = 10 misses U(B) = 0 misses

Problem: GA considers benefit only from the imme-diate block. Hence it fails to exploit huge gains from ahead

Blocks assigned

Mis

ses

All blocks assigned to A, even if B has same miss reduc-tion with fewer blocks

Lookahead Algorithm

Marginal Utility (MU) = Utility per cache resource MUa

b = Uab/(b-a)

GA considers MU for 1 block. LA considers MU for all possible allocations

Select the app that has the max value for MU. Allocate it as many blocks required to get max MU

Repeat till all blocks assigned

Lookahead Algorithm (ex-ample)

Time complexity ≈ ways2/2 (512 ops for 32-ways)

0

10

20

30

40

50

60

70

80

90

100

0 1 2 3 4 5 6 7 8

A

B

Iteration 1:MU(A) = 10/1 block MU(B) = 80/3 blocks

B gets 3 blocks

Result: A gets 5 blocks and B gets 3 blocks (Op-timal)

Next five iterations: MU(A) = 10/1 block MU(B) = 0A gets 1 block

Blocks assigned

Mis

ses

Results for partitioning al-gorithms

Four cores sharing a 2MB 32-way L2

Mix2(swm-glg-mesa-prl)

Mix3(mcf-applu-art-vrtx)

Mix4(mcf-art-eqk-wupw)

Mix1(gap-applu-apsi-

gzp)

LA performs similar to EvalAll, with low time-com-plexity

LRUUCP(Greedy)UCP(Lookahead)UCP(EvalAll)

Summary

CMP and shared caches are common

Partition shared caches based on utility, not demand

UMON estimates utility at runtime with low overhead

UCP improves performance: Weighted speedup by 11% Throughput by 17% Hmean-fairness by 11%

Lookahead algorithm is scalable to many cores shar-ing a highly associative cache