TurboROB A Low Cost Checkpoint/Restore Accelerator

1
TurboROBA Low Cost Checkpoint/Restore Accelerator
Patrick Akl and Andreas Moshovos AENAO Research
Group Department of Electrical and Computer
Engineering University of Toronto pakl,
moshovos_at_eecg.toronto.edu
2
Recovering From Control Flow Mispredictions

• Execution Timeline

Predict a Branch Outcome
Misprediction Discovered
Recover Processor State Redirect Fetch
Correct Path
Predicted Path
Resume Execution

• Accelerate Recovery Improve Performance

3
State-of-the-Art Recovery
Log of Changes
State Snapshot
Predict a Branch Outcome
ROB
Misprediction Discovered
what
old value

• Scalability and/or Performance Issues

4
Turbo-ROB
Log of Changes
Predict a Branch Outcome
ROB
Misprediction Discovered
Partial Log of Changes

• Make common case fast
• Recover only at branches
• Store only as much as needed
• Partial Log

5
Outline

• Control Flow Mispeculation Recovery
• TurboROB
• Methodology and Results
• Summary

6
State Recovery Example Register Alias Table
Lg( arch. regs)
Original Code
RAT
A add r1, r2, 100 B breq r1, E C sub r1, r2, r2
p1
p4
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
7
ROB Slow, Fine-Grain Recovery

• Each entry contains
• Architectural destination register
• Its previous RAT map

Program Order
3. Undo RAT updates in reverse order
Reorder Buffer

• Misprediction discovered

• 2. Locate newest instruction

INVALID
RAT

• Too slow recovery latency proportional to number
  of instructions to squash

8
Global Checkpoints Fast, Coarse-Grain Recovery
Program Order
checkpoint
checkpoint
checkpoint
checkpoint
Reorder Buffer

• Misprediction discovered

INVALID
RAT

• Branch w/ GC Recovery is Instantaneous

9
Impact of More Checkpoints
Concept
Actual Implementation
architectural register
physical register

• More checkpoints ?
• Power hungry structure
• Increased delay
• Only a few checkpoints can practically be
  implemented
• Cannot always cover all branches

10
Intelligent Checkpointing BranchTap
checkpoint
checkpoint
checkpoint
checkpoint

• Use Few Checkpoints Effectively
• BranchTap
• Throttle Speculation

11
Conventional Mechanisms Recovery Scenarios
B
B
B
checkpoint
B
B
B
checkpoint
Re-Execution
B
B
B
checkpoint
12
Outline

• Background
• Turbo-ROB
• Methodology and Results
• Summary

13
Turbo-ROB
Recovery Cost
B
R2
R1
R1
R1
R2
ROB Recovery
useful
redundant
We only need to reverse the first subsequent
change for every RAT entry
14
Turbo-ROB Replacing the ROB
B
B
B
TROB
Re-Execution
B
B
B
TROB
15
Selective Turbo-ROB w/ ROB
B
B
B
TROB
Selective Turbo-ROB w/ GCs
B
B
B
TROB
checkpoint
16
Outline

• Background
• TurboROB
• Methodology and Results
• Summary

17
Results Overview

• TROB as an ROB replacement
• BranchTap offers better performance than ROB
• Fewer resources
• Even for smaller windows
• Selective TROB as a GC reduction mechanism
• TROB reduces pressure for GCs
• Offload a critical structure RAT
• In the paper
• Selective TROB as an ROB accelerator
• Even the smallest TROB accelerates recovery

18
Methodology

• Simulator based on Simplescalar
• Alpha/OSF
• 24 SPEC CPU 2000 benchmarks
• Reference Inputs
• Processor configurations
• 4-way OoO core
• 128/256/512 in-flight instructions
• 1K-entry confidence table for low confidence
  branch identification / similar results with
  Anyweak
• 1B committed instructions after skipping 2B

19
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

20
TROB Replacing the ROB/512-Entry Window

• 64-entry TROB ROB on the Average
• Pathological cases exist ? 256-entry needed
• 512-Entry TROB better than ROB

21
TROB Replacing the ROB/128-Entry Window

• 64-Entry 50 better than ROB
• Fewer pathological cases
• 128-Entry TROB better than ROB

22
sTROB and Global Checkpoints/128-Entry Window

• TROB 1 GC better than 4GCs

23
Summary

• TROB vs. ROB
• Replacement
• Same resources ? better performance
• Fewer resources ? often better performance
• Except when accuracy is high
• Acceleration
• ¼ resources ? 35 improvement
• TROB vs. GCs
• Reduce pressure from the critical path
• With just 1 GC match the performance of four GCs
• One more alternative for designers
• Allows different area/performance/power tradeoffs

24
TurboROBA Low Cost Checkpoint/Restore Accelerator
Patrick Akl and Andreas Moshovos AENAO Research
Group Department of Electrical and Computer
Engineering University of Toronto pakl,
moshovos_at_eecg.toronto.edu
25
TROB Replacing the ROB/512-Entry Window

• 64-entry TROB ROB on the Average
• Pathological cases exist ? 256-entry needed
• 512-Entry TROB better than ROB

26
TROB Replacing the ROB/128-Entry Window

• 64-Entry 50 better than ROB
• Fewer pathological cases
• 128-Entry TROB better than ROB