SMARTS:Accelerating Microarchitecture Simulation via Rigorous Statistical Sampling

1
SMARTSAccelerating Microarchitecture Simulation
via Rigorous Statistical Sampling

• Roland E. Wunderlich Thomas F. Wenisch Babak
  Falsafi James C. Hoe
• Computer Architecture Laboratory
• Carnegie Mellon University, Pittsburgh, PA
• Joo hyung Kim

2
Contents

• Introduction
• Statistical sampling
• The SMARTS framework
• SMARTS in practice
• Using SMARTS
• Conclusion

3
Introduction

• Current approaches
• Abbreviated instruction execution streams
• Fewer or smaller input sets
• Shortcomings
• On the efficiency front, large sampling units
• On the accuracy front, no tight error bounds
• The SMARTS approach
• Statistical sampling theory
• Exact and constructive procedure

4
Statistical sampling

• Theory of sampling
• Choosing a minimal
• Representative sample to achieve a quantifiable
  accuracy and precision in the estimate
• Not presume a normally distributed population
• Our goal
• Identify a minimal but representative sample from
  the population for microarchitecture simulation
• Establish a confidence level for the error on
  sample estimates

Sampling variables
5
The SMARTS framework

• Technique overview
• Detailed mode
• Functional mode
• Detailed warming short-comings
• Expensive
• Difficult to derive analytically
• Functional warming
• The cache hierarchies and branch predictors are
  prime candidates

SMARTS variables
Systematic sampling in SMARTS
6
The SMARTS framework

• Benchmarks
• Demonstrate the effectiveness of SMARTS
• CPI and EPI of the SPEC CPU2000 integer and
  floating-point
• SimpleScalar 3.0sim-outorder simulator

Machine configurations
7
The SMARTS framework

• Speedup opportunity

Coefficient of variation of CPI
VCPI decreases with increasing U because
Short-term CPI variations within a window of U
instructions are hidden by averaging over the
sampling unit.
8
The SMARTS framework

• Simulation speedup model

Modeled SMARTS simulation rate The two SD plots
show the simulation rate without function
warming. The SFW plot shows the simulation rate
when using functional warming to bound W.
9
SMARTS in practice

• SMARTSim
• Sim-outorder
• Support functional simulation prior to starting
  detailed simulation
• SMARTSim
• Repeated transitions back-and-forth between
  functional and detailed simulation modes
• Accepts the systematic sampling parameters
• Fast-forwarding options
• Functional simulation
• Functional simulation with warming

10
SMARTS in practice

• Optimal sampling unit size

Optimal U
The left chart shows that the optimal U increases
with W. The right chart shows that U 1000 is a
reasonable choice across benchmarks and W.
11
SMARTS in practice

• Effectiveness of detailed warming

Detailed warming requirements without functional
warming. (8-way) Choosing U 1000 and n
sufficient for a 99.7 confidence interval of 3
12
SMARTS in practice

• Effectiveness of functional warming
• All benchmarks have bias under 2.0
• Only 6 benchmarks in each configuration exceed
  1.0

CPI bias achieved with functional warming and
minimal detailed warming
13
Using SMARTS

• SMARTS procedure
• W is selected to exceed the bounded history of
  the microarchitectural state
• Setting U 1000
• Determine n, and correspondingly k
• Correct value for n
• A sampling measurement is made using a generic
  initial value ninit
• ntuned for a second run is calculated from the Vx
  of the initial run

14
Using SMARTS

• Comparison to SimPoint
• SimPoint advantages
• Obviates the need for functional warming
• Quick integration into a simulation
  infrastructure
• Early termination of simulation
• SimPoint shortcomings
• High CPI error
• Unquantifiable confidence in estimates

15
Using SMARTS

• Comparison to SimPoint
• SimPoint has a higher average error (3.7 vs.
  0.6)
• higher worst-case error (-14.3 for gcc-2)
• SimPoint estimate based on just a single instance
  of the basic block sequences yields a large
  error.
• SMARTS uses the measured coefficient of variation
  to help gauge both the required sample size and
  the confidence in the estimates.

Comparison of SMARTS with SimPoint SimPoints
mean runtime per benchmark is 2.8 hours compared
to 5.0 hours for SMARTS.
16
Conclusion

• Evaluation Results
• Average error 0.64 on CPI and 0.59 on EPI
• Average speedups of 35 and 60 times
• Future simulator designs
• Designers should focus on the simulators
  flexibility and realism
• Designers should focus on techniques to speed up
  fast-forwarding and functional warming