BranchTap Improving Performance With Very Few Checkpoints Through Adaptive Speculation Control Patrick Akl and Andreas Moshovos AENAO Research Group Department of Electrical and Computer Engineering University of Toronto
What Happens on a Branch Misprediction? Execution Timeline Predict a Branch Outcome Predicted Path Correct Path Misprediction Discovered Recover Processor State Redirect Fetch Resume Execution We wish to make the recovery fast
State-of-the-art recovery Existing mechanisms Reorder buffer based: slow Instantaneous checkpoints: faster Problem: can’t have enough checkpoints State-of-the-art solution: checkpoint prediction Allocate the few checkpoints judiciously Another degree of freedom: speculation control Sometimes deeper speculation = higher recovery cost Can hurt performance Throttle speculation
BranchTap Results / Benefits No additional checkpoints are needed Dynamically adapts to application behavior Improves performance for most programs Misprediction performance penalty reduced by 28% on AVG BranchTap comes “for free” Very simple to implement Better than more accurate checkpoint predictors
Outline Background BranchTap Methodology and Results Summary
State Recovery Example: Register Alias Table Original Code Lg(# arch. regs) RAT A add r1, r2, 100 B breq r1, E C sub r1, r2, r2 p4 p1 p5 p5 p4 Architectural Register p2 p3 # arch. regs Renamed Code A add p4, p2, 100 B breq p4, E C sub r5, p2, p2 Physical Register
ROB: Slow, Fine-Grain Recovery Each entry contains Architectural destination register Its previous RAT map Program Order 3. Undo RAT updates in reverse order B B B B B Reorder Buffer Misprediction discovered 2. Locate newest instruction INVALID RAT Too slow: recovery latency proportional to number of instructions to squash
Global Checkpoints: Fast, Coarse-Grain Recovery Program Order checkpoint checkpoint checkpoint checkpoint B B B B B Reorder Buffer Misprediction discovered INVALID RAT Branch w/ GC: Recovery is “Instantaneous”
Impact of More Checkpoints Concept Actual Implementation Working Copy checkpoints RAT architectural register physical register More checkpoints ? Power hungry structure Increased delay Only a few checkpoints can practically be implemented Cannot always cover all branches
Intelligent Checkpointing State of the art solution Checkpoint allocation: Allocate checkpoints at hard-to-predict branches Checkpoint management: Release checkpoints as soon as they are no longer needed Use few checkpoints efficiently
Conventional Mechanisms: Recovery Scenarios Mispeculation on a branch w/ a GC: Direct recovery Mispeculation on a branch w/o a GC: Indirect recovery With intelligent checkpointing: 30% Indirect recoveries 75% of performance loss B B B ROB Fast Recovery checkpoint B B B ROB Slow Recovery checkpoint
Outline Background BranchTap Methodology and Results Summary
BranchTap Motivation Low confidence branch ~ Recovery Cost No Wait Scenario B B B ROB checkpoint checkpoint Misprediction discovered Wait Scenario B B B ROB ~ Recovery Cost checkpoint checkpoint Sometimes, it is better to wait if no checkpoint is available
BranchTap Concept Key idea: stall when speculation is likely to deteriorate performance Count the number of low confidence branches w/o a checkpoint If it exceeds a threshold, stall Threshold selection Fixed Varies greatly across programs Can deteriorate performance significantly Adaptive Robust performance Minimize recovery cost while conserving good speculation opportunities
Threshold Adaptation Policy BranchTap adapts across and within applications
Outline Background BranchTap Methodology and Results Summary
Results Overview Performance w/o Checkpoints BranchTap improves even with just an ROB Performance w/ 4 Checkpoints BranchTap improves over conventional recovery methods Performance w/ Larger Checkpoint Predictors BranchTap offers better performance than a 64x larger predictor
Methodology Simulator based on Simplescalar 24 SPEC CPU 2000 benchmarks Reference Inputs Processor configurations 8-way OoO core Up to 1K in-flight instructions 1K-entry confidence table for low confidence branch identification 1B committed instructions after skipping 100B
“Perfect Checkpointing” Configuration A checkpoint is auto-magically taken at all mispredicted branches All recoveries are fast We report the “deterioration relative to perfect checkpointing” We compare BranchTap against the obvious solution of unrestricted speculation. We normalize our performance results relative to a “Perfect Checkpointing” configuration where we assume all mispredictions are automagically checkpointed.
Performance with No Checkpoints Deterioration relative to “perfect checkpointing” better -39% deterioration BranchTap improves over conventional mechanisms Adaptation leads to robust performance improvements
Performance Evaluation with 4 Checkpoints Deterioration relative to “perfect checkpointing” BranchTap with 4 checkpoints is better than 6 checkpoints alone better -28% deterioration
BranchTap vs. Larger Checkpoint Predictors BranchTap with a 1K-entry confidence table and 4 GCs: Higher performance than a 64K-entry confidence table with 4 GCs Lower complexity, virtually comes “for free” better deterioration BranchTap confidence table size
Outline Background BranchTap Methodology and Results Summary
Summary Performance with 4 (no) checkpoints ~28 (39) % of misprediction penalty removed BranchTap is robust: Up to 6 (13) % better and max 1.2 (0.1) % worse than conventional mechanisms BranchTap is very simple to implement Few counters and comparators BranchTap is better than other alternatives BT + 1K predictor better than a 64K predictor alone BT + 4 GCs better than 6 GCs alone
BranchTap Improving Performance With Very Few Checkpoints Through Adaptive Speculation Control Patrick Akl and Andreas Moshovos AENAO Research Group Department of Electrical and Computer Engineering University of Toronto {pakl, moshovos}@eecg.toronto.edu