Understanding the Backward Slices of Performance Degrading Instructions

Size: px

Start display at page:

Download "Understanding the Backward Slices of Performance Degrading Instructions"

Branden Curtis
6 years ago
Views:

1 Understanding the Backward Slices of Performance Degrading Instructions Craig Zilles and Guri Sohi University of Wisconsin - Madison International Symposium on Computer Architecture June, 2000

2 Motivation PRCESSRS ACHIEVE NLY A FRACTIN F PEAK PERFRMANCE N MANY PRGRAMS Performance Degrading Events (PDE) branch mispredictions cache misses RETIREMENT STREAM BRANCH branch mispredict TIME LARGER CACHES AND PREDICTRS handle easy cases concentrates PDE s to a fraction of problem static instructions UNCRRELATED (DATA-DEPENDENT) BRANCHES HASH TABLE LKUPS AND PINTER CHASING cache miss LAD International Symposium on Computer Architecture (ISCA-2000), June

3 Motivation, cont. PRGRAM BEHAVIR IS DETERMINISTIC build predictors which use the program! Pre-execution FRK Pre-executed sub-program feeds prediction for branch fetched by the main thread sub program Pre-fetch memory similarly BRANCH BRANCH UTCME BRANCH TIME nly pre-execute instructions which defy normal predictors AVID MISPREDICTIN International Symposium on Computer Architecture (ISCA-2000), June

4 Motivation, cont. THE EFFECTIVENESS F PRE-EXECUTIN IS DETERMINED BY THE SUB-PRGRAM sub program BRANCH Sub-program must enable faster execution of the problem instruction Sub-program size determines overhead BRANCH How can we build sub-programs to minimize their size while maintaining accuracy International Symposium on Computer Architecture (ISCA-2000), June

5 verview MTIVATIN PRGRAM SLICING EXPERIMENT VERVIEW METHDLGY CNSERVATIVE SLICES EXAMPLE SPECULATIVE PTIMIZATINS ADDITINAL BSERVATINS CNCLUSINS International Symposium on Computer Architecture (ISCA-2000), June

6 Program Slicing THE SUB-PRGRAM INCLUDES NLY THE SUBSET F INSTRUCTINS WHICH CAN INFLUENCE THE PRBLEM INSTRUCTIN. lda cmoveq ldl s4addq stl and sll ldq ldl addl addq bis stl ldq r8,-8432(r29) r18,r1,r0 r1,-19952(r29) r16,r8,r8 r31,0(r8) r4,r5,r5 r5,4,r5 r23,-19408(r29) r27,-19944(r29) r1,1,r1 r0,r5,r5 r31,r31,r0 r1,-19952(r29) r7,8(r5) BACKWARD SLICE cmoveq and sll addq ldq r18,r1,r0 r4,r5,r5 r5,4,r5 r0,r5,r5 r7,8(r5) follow dependences backward from criterion instruction both data and control Criterion Instruction International Symposium on Computer Architecture (ISCA-2000), June

7 Experiment verview INITIAL CHARACTERIZATIN F SLICES Categorization of instructions in the slice largest contributors: control and memory dependence resolution NT dataflow Exploiting well-known phenomena to reduce slice size highly-biased branches stability of memory dependences International Symposium on Computer Architecture (ISCA-2000), June

8 Methodology SPEC95 integer benchmarks Alpha architecture, optimized -4 Profiled to identified problem static instructions Frequently caused mispredictions or cache misses cum. slice size Generated ASSEMBLY-LEVEL slices Looked at the 512 dynamic instructions leading to the criterion Removed NPS, and SP/GP computation dependences all instructions cumulative slice size distance (in dynamic instructions) International Symposium on Computer Architecture (ISCA-2000), June

9 Conservative Slices PRGRAMS HAVE AMBIGUUS MEMRY DEPENDENCES AND CMPLEX CNTRL FLW CNSERVATIVELY CNSTRUCTED SLICES CAN BE LARGE 50% of program necessary to compute all store addresses 80% of program necessary to resolve all branches SLUTIN: EXPLIT THE FACT THAT SLICES NLY PRVIDE HINTS construct speculative slices assume common-case behavior profiling is required to detect the common-case TW EXAMPLE PTIMIZATINS: both targeting ambiguous memory dependences International Symposium on Computer Architecture (ISCA-2000), June

10 Speculative ptimizations A CNSERVATIVE SLICE MUST CMPUTE THE ADDRESS FR EVERY STRE WHICH CULD ALIAS WITH A LAD IN THE SLICE THE SET F MEMRY DEPENDENCES ACTUALLY REALIZED IS A SUB-SET F THSE WHICH ARE PSSIBLE cum. slice size Profile to identify the store sets nly compute store addresses for these stores conservative (all aliasable stores) all instructions store set oracle distance (in dynamic instructions) Slices built using store sets approximate size of oracle slices International Symposium on Computer Architecture (ISCA-2000), June

11 Speculative ptimizations, cont. WHEN MEMRY DEPENDENCES EXIST, FTEN THE LAD CMMUNICATES WITH THE MST RECENT STRE FRM ITS STRE SET. much like communication through registers EXPLIT T REDUCE SLICE SIZE: 1. Assume a communication pattern (imprecise transformation) 2. Remove load and store from the slice STRE R9 -> 0(R17) ADD R10, R11 -> R12 STRE R12 -> 0(R18) LAD 0(R19) -> R13 ADD R13, R14 -> R15... STRE R9 -> 0(R17) ADD R10, R11 -> R12 STRE R12 -> 0(R18) LAD 0(R19) -> R13 ADD R12, R14 -> R15... International Symposium on Computer Architecture (ISCA-2000), June

12 Speculative ptimizations, cont. STATIC LADS ARE HIGHLY BIASED WITH RESPECT T THIS BEHAVIR Mis-speculation can be avoided Easily profiled to classify the dependences cum. slice size all instructions all aliasable stores store set oracle optimized distance (in dynamic instructions) Removing such load-store pairs can further reduce slice size with little affect on accuracy Reduced address sub-slice to 1/4 of conservative size International Symposium on Computer Architecture (ISCA-2000), June

13 Additional bservations FTEN DATA DEPENDENCES ARE CLUSTERED NEAR CRITERIN possibly influences pre-execution mechanisms NT UNCMMN FR SLICES T VERLAP create a single multi-slice SLICES ARE BURSTY due to program structure cum. slice size Regions with no contribution distance (in dynamic instructions) International Symposium on Computer Architecture (ISCA-2000), June

14 Additional bservations FALSE CNTRL DEPENDENCES: control dependent regions are part of slice, but all paths from the branch contribute to the slice equivalently. CMMN CASE: CNDITINAL FUNCTIN CALL CRITERIN INT A =...; IF (B) { FUNCTIN(); } IF (A) FUNCTIN() { SAVE A;... RESTRE A; } THER CASE: CDE REPLICATIN Currently refining infrastructure to handle these cases International Symposium on Computer Architecture (ISCA-2000), June

15 Summary PRE-EXECUTIN: GENERAL TECHNIQUE FR HANDLING PRBLEM INSTRUCTINS Use the program to predict the program Requires small, accurate slices SPECULATIVE SLICES: Exploit common-case behavior to reduce slice size while maintaining accuracy Some slices can be reduced to less than 10% of the 512 instructions preceding the criterion while maintaining greater than 95% accuracy Much future work to be done Requires sophisticated profile information International Symposium on Computer Architecture (ISCA-2000), June

Speculative Multithreaded Processors

Speculative Multithreaded Processors Guri Sohi and Amir Roth Computer Sciences Department University of Wisconsin-Madison utline Trends and their implications Workloads for future processors Program parallelization and speculative threads