Performance Evaluation of Adaptivity in STM. Mathias Payer and Thomas R. Gross Department of Computer Science, ETH Zürich

Transcription

1 Performance Evaluation of Adaptivity in STM Mathias Payer and Thomas R. Gross Department of Computer Science, ETH Zürich

2 Motivation STM systems rely on many assumptions Often contradicting for different programs Statically tuned to a baseline Use self-optimizing systems Adapt to different workloads What parameters can be adapted? How to measure effectiveness? ISPASS'11 / Mathias Payer / ETH Zürich 2

3 Outline Introduction STM System STM Baseline Adaptive Parameters Evaluation Related work Conclusion ISPASS'11 / Mathias Payer / ETH Zürich 3

4 Introduction Software Transactional Memory (STM) applies transactions to memory (Optimistic) concurrency control mechanism Alternative to lock-based synchronization Multiple concurrent threads run transactions Concurrent memory modifications ISPASS'11 / Mathias Payer / ETH Zürich 4

5 Introduction Concurrent transactions modify memory without synchronization Transaction is verified after completion Conflicts are detected and resolved Changes committed for conflict-free transactions Modifications only visible after commit ISPASS'11 / Mathias Payer / ETH Zürich 5

6 withdraw { tmp = balance; tmp = tmp 1 balance = tmp; } Introduction deposit { tmp = balance; tmp = tmp + 1 balance = tmp; } TX starts balance in read-set balance in write-set Conflict detection, data committed What happens when balance is accessed concurrently? Either locking or STM needed to ensure correct end balance STM system decides which tx is executed first ISPASS'11 / Mathias Payer / ETH Zürich 6

7 STM Baseline Many efficient STM implementations agree on important design decisions: Word-based locking Global locking / version table Eager locking (Almost) no contention management Simple write-set and read-set implementations ISPASS'11 / Mathias Payer / ETH Zürich 7

8 STM Baseline Combined global write lock / version array Write list / buffer Lock list Read list / buffer Write list / buffer Lock list Read list / buffer Write Hash Read Hash Write Hash Read Hash Transaction Transaction ISPASS'11 / Mathias Payer / ETH Zürich 8

9 Adaptive STM Parameters Global adaptivity Synchronization needed Optimizes to global optimum Averages over all concurrent transactions (Thread-) local adaptivity No synchronization needed Limits adaptable parameters Best parameters for each thread/transaction ISPASS'11 / Mathias Payer / ETH Zürich 9

10 Adaptive STM Parameters Different adaptive parameters measured: Size of global locking/version-table *G Size of local hash-tables *L Write strategy *L Locality tuning for hash-functions *L Contention management *L *L local, *G global ISPASS'11 / Mathias Payer / ETH Zürich 1

11 Adaptive Hash-Table Global hash-table: trade-off between overlocking and locality Global strategy: coordinate lock collisions and overlocking between threads Adapt size based on global information Local hash-table: trade-off between reset cost, and # hash-collisions Local strategy: sample moving average of unique write locations Adapt size based on trend ISPASS'11 / Mathias Payer / ETH Zürich 11

12 Adaptive Write Strategy Different costs depending on strategy Write-back: cheap abort, expensive commit Write-through: expensive abort, cheap commit Adapt strategy to per-thread workload Measure abort rate ISPASS'11 / Mathias Payer / ETH Zürich 12

13 Adaptive Locality Tuning Different applications have different data access patterns No optimal hash function for all data accesses Measure number of hash collisions for threadlocal hash tables Circle through different hash functions ISPASS'11 / Mathias Payer / ETH Zürich 13

14 Adaptive Contention Management No single strategy works in all environments Measure contention and implement an adaptive back-off strategy Wait and retry Abort later ISPASS'11 / Mathias Payer / ETH Zürich 14

15 Local Adaptive STM Parameters (for local hash-table) # writes vs. hash-table space enlarge write-hash no change shrink write-hash ISPASS'11 / Mathias Payer / ETH Zürich 15

16 Local Adaptive STM Parameters (for local hash-table) no change change hash-function # hash collisions ISPASS'11 / Mathias Payer / ETH Zürich 16

17 Local Adaptive STM Parameters (for local hash-table) # writes vs. hash-table space enlarge write-hash no change shrink write-hash # hash collisions enlarge write-hash & change hash-function change hash-function shrink write-hash & change hash-function ISPASS'11 / Mathias Payer / ETH Zürich 17

18 AdaptSTM Adaptive STM system built on presented features Statically tuned competitive baseline Static global hash function and hash table Mature and stable implementation Different local adaptive parameters Write-set hash function and size of hash table Write-through and write-back write strategy Adaptive contention management ISPASS'11 / Mathias Payer / ETH Zürich 18

19 Evaluation Benchmark: STAMP configuration (increased workload for kmeans) AdaptSTM version.5.1 Intel 4-core Xeon E552 CPU GHz, 12GB RAM 64bit Ubuntu 9.4 ISPASS'11 / Mathias Payer / ETH Zürich 19

20 Evaluation: Global Hash-Table kmeans 4 Threads Genome 4 Threads Time [s] ^16 2^18 2^2 2^22 2^24 2^26 Time [s] ^16 2^18 2^2 2^22 2^24 2^ # Shifts # Shifts ISPASS'11 / Mathias Payer / ETH Zürich 2

21 Evaluation: Global Adaptivity Global optimizations have limited potential Small optimization potential High synchronization cost Reasonable baseline outperforms global optimization ISPASS'11 / Mathias Payer / ETH Zürich 21

22 Evaluation: Local Adaptivity Different configurations: nawb: no adaptivity, use write-back awbt: adaptivity, adjust write-through / write-back awwh: awbt plus an adaptive hash-table for the write-set awhh: awwh plus different hash functions aall: all adaptive parameters plus Bloom filter for write-entries Adaptation system starts with best 'average' parameters, improves from there ISPASS'11 / Mathias Payer / ETH Zürich 22

23 Evaluation: Local Adaptivity kmeans Labyrinth 15.% 3.% 1.% 2.% Speedup to non adaptive 5.%.% -5.% Speedup to non adaptive 1.%.% awbt awwh awhh -1.% aall -2.% awbt awwh awhh aall -1.% -3.% -15.% % Threads Threads awbt: adaptive, write-back/-through awwh: adaptive, write-back/-through, write-hash awhh: adaptive, write-back/-through, write-hash, hash-function aall: adaptive, write-back/-through, write-hash, hash-function, Bloom filter ISPASS'11 / Mathias Payer / ETH Zürich 23

24 Evaluation: Local Adaptivity 6.% Genome 5.% Vacation Speedup to non adaptive 5.% 4.% 3.% 2.% 1.%.% -1.% Speedup to non adaptive 4.% 3.% awbt 2.% awwh awhh 1.% aall.% awbt awwh awhh aall -2.% -1.% -3.% Threads -2.% Threads awbt: adaptive, write-back/-through awwh: adaptive, write-back/-through, write-hash awhh: adaptive, write-back/-through, write-hash, hash-function aall: adaptive, write-back/-through, write-hash, hash-function, Bloom filter ISPASS'11 / Mathias Payer / ETH Zürich 24

25 Evaluation: Local Adaptivity No single optimization works for all benchmarks Combination of all options leads to best performance Impressive speed-ups for individual benchmarks compared to the globally optimized case ISPASS'11 / Mathias Payer / ETH Zürich 25

26 Related Work TL2 (Dice et al.): baseline STM system Different related work on static tuning of global parameters (Harris, Dice, Ennals, Felber) Crucial for efficient baseline TinySTM (Felber et al.): adapts size and hash function of global locking table ASTM (Marathe et. al.): adapts lazy-eager locking strategies and different meta-formats ISPASS'11 / Mathias Payer / ETH Zürich 26

27 Conclusions Adaptivity in STM is important for good performance Speedups up to 1% possible Global optimization are limited Low potential, high synchronization cost Local optimizations tune thread-local parameters High correlation with workload ISPASS'11 / Mathias Payer / ETH Zürich 27

28 Questions? Contact: Source: ISPASS'11 / Mathias Payer / ETH Zürich 28

29 Evaluation: Global Hash-Table Bayes Genome 4 Threads 4 Threads Time [s] Time [s] ^16 2^18 2^2 2^22 2^24 2^ # Shifts # Shifts Vacation kmeans 4 Threads 4 Threads 3 8 Time [s] Time [s] ^16 2^18 2^2 2^22 2^24 2^ # Shifts # Shifts ISPASS'11 / Mathias Payer / ETH Zürich 29

30 Evaluation: Global Hash-Table Labyrinth Intruder 4 Threads 4 Threads Time [s] ^16 2^18 2^2 2^22 2^24 2^26 Time [s] ^16 2^18 2^2 2^22 2^24 2^ # Shifts # Shifts SSCA2 YADA 4 Threads 4 Threads 18 5 Time [s] ^16 2^18 2^2 2^22 2^24 2^26 Time [s] ^16 2^18 2^2 2^22 2^24 2^ # Shifts # Shifts ISPASS'11 / Mathias Payer / ETH Zürich 3

31 STM Comparison 1.6 Genome 5 Vacation Relative runtime Relative runtime astm 2.5 tl2 tstm2 tstm astm tl2 tstm tstm Threads Threads 1.8 Labyrinth 6 Intruder Relative runtime Relative runtime 4 astm 3 tl2 tstm tstm99 2 astm tl2 tstm tstm Threads Threads ISPASS'11 / Mathias Payer / ETH Zürich 31

32 Evaluation: Local Adaptivity Bayes SSCA2 3.% 5.% 2.% 4.% 1.% 3.% Speedup to non adaptive.% -1.% -2.% Speedup to non adaptive 2.% awbt awwh 1.% awhh aall.% -1.% awbt awwh awhh aall -3.% -2.% -4.% Threads -3.% Threads ISPASS'11 / Mathias Payer / ETH Zürich 32

33 Evaluation: Local Adaptivity 12.% YADA 1.% 8.% Speedup to non adaptive 6.% 4.% 2.% awbt awwh awhh aall.% -2.% Threads ISPASS'11 / Mathias Payer / ETH Zürich 33