Aashish Phansalkar & Lizy K. John

1

Aashish Phansalkar&

Lizy K. John

Performance Prediction Using Program Similarity

The University of Texas at AustinThe University of Texas at Austin

2

Outline Motivation and Objectives

Methodology

Experimental results

Conclusion

Future work

3

Motivation (1): Simulation is costly

A computer architect or a designer has to simulate multiple customer applications

Simulations take very long due to the complexity of modern microprocessor designs

4

Motivation(2): Making a decision based on benchmark

scores

Customers often use benchmarks to make a decision about buying computer systems

The application program they use often, may not be a part of the benchmark suite

Customers can use benchmarks as representatives of their application programs

Predict performance of their application based on the already available performance data of benchmarks

5

Objective

A quantitative method to estimate performance without running cycle accurate simulation

Use the knowledge of similarity between a customer’s application program and known benchmark programs to develop a quantitative approach to predict performance

6

Outline Motivation and Objectives

Methodology

Experimental results

Conclusion

Future work

7

Overview

Customer application

Repository of

Benchmarks

Measure similarity

Predicted performance

New CaseKnown cases

8

Program characterization Instruction mix

Percentage of different types of instructions e.g. percentage of memory references, percentage of branch

instructions

Control Flow % Taken branches % Forward branches % Forward taken branches Basic Block Size (Number of instructions between two branches)

Register Dependency Distance Data and instruction temporal locality of program

Data and instruction spatial locality of program

9

ADD R1, R3,R4MUL R5,R3,R2ADD R5,R3,R6LD R4, (R8)SUB R8,R2,R1

Read After Write Dependency Distance = 4

Measure Distribution of % of dependency distances for following set of ranges.

1, 2, 3-4, 5-8, 8-16, 16-32, greater than 32

Normalized count for each range of dependency distance forms a metric

Register dependency distance

10

Memory reuse distance 2004, 2008, 4008, 2000, 1080,2004,4008

Reuse Distance = 4 Reuse Distance = 3

Data and instruction temporal locality

Computing reuse distance for a trace of byte addresses is very computation and space intensive

Reuse distance for a block of 16, 64, 256, 4096 bytes

Temporal locality metrics (tlocality) = Wt. average reuse distance

11

Data and instruction spatial locality Spatial locality metrics are derived from the temporal locality metrics

As the block size increases, programs with good spatial locality will show lower values for tlocality for higher block sizes

Spatial Locality = tlocality64 / tlocality16 tlocality256 / tlocality16

tlocality4096 / tlocality16

12

Methodology Overview

Microarchitecture independent metrics for known benchmarks

Microarchitecture independent metrics for the customer application

Measure program similarity

Prediction of target metric

for new application(2 methods)

Similarity information

Predicted value of target metric

13

Measuring Similarity (1)

Distance between two programs in the workload space is the measure of their similarity

We assume that similarity between two programs is inversely proportional to the Euclidean distance between them

14

Measuring similarity (2)

The workload space is made of many workload characteristics and so its dimensionality is very high

Inherent characteristics are highly correlated Euclidean distance measured using these characteristics will be

biased The correlated variables will add twice to the distance as the

independent variables

Use Principal Components Analysis (PCA)

15

Method 1: Predicting performance using weights

Compute distance of similarity from program X to each benchmark program dx1, dx2, dx3…dxn in the PC space

Calculate weights w1, w2, ….

w1

w2

User program X

benchmarks

16

Method 2: Predicting performance using clustering

Measure all the inherent characteristics for the benchmarks and user program X

Cluster all the programs based on the inherent characteristics and find optimal clusters

User program X

benchmarks

17

Outline Motivation and Objectives

Methodology

Experimental results

Conclusion

Future work

18

Experiments Used integer programs from SPEC CPU2000 suite

to demonstrate the use of Method 1 and Method 2 described

Prediction of speedup Used all the workload characteristics to form the workload space

Prediction of cache miss-rates Used only the data locality characteristics to form the workload space

19

Experiment: Predict performance (speedup) of bzip2

using benchmarks from SPEC CPU2000 suite Assume that bzip2 is the customer application

Performance of SPEC CPU2000 benchmarks is known

Predicting speedup(1)

 SPEC int 2000 benchmarks used for prediction Speedup for SGI Altix (1500MHz, Itanium 2)

164.gzip 9.09

197.parser 7.7

300.twolf 11.9

176.gcc 12.48

252.eon 12.52

186.crafty 10.46

255.vortex 15.54

175.vpr 10.39

181.mcf 13.01

Speedup for each benchmark program on a machine (from the scores reported on the SPEC website)

20

Predicting speedup(2)

 Mean used % Error in predicting speedup

Wt GM Error 4.69

Wt HM Error 2.5

Wt AM Error 6.87

GM Error 8.68

HM Error 6.53

AM error 10.77

Clustering 12.08

Method 1:Predicting speedup using weightsMachine name: SGI-Altix 3000 (1500MHz, Itanium 2)

21

Predicting speedup (3)

Statistics

% Error in predicted speedup

usingweighted GM

% Error in predicted speedup

using weighted HM

% Error in predicted speedup

using weighted AM

% Error in predicted speedup

using GM

Average 14.37 11.31 18.05 16.41

Lower CI(95%) 13.72 10.74 17.30 15.73

Upper CI(95%) 15.02 11.89 18.80 17.09

Method 1: Predicting speedup using weightsFor 50 different machines the error in predicted speedup

22

Predicting speedup (4)

Cluster 1 parser, twolf, vortex

Cluster 2 bzip2, gzip

Cluster 3 eon, vpr

Cluster 4 mcf

Cluster 5 crafty

Cluster 6 gcc

Method 2: Predicting speedup using clustering

The average error in predicting the speedup over all machines for bzip2 is 20.29%

23

Prediction of data cache miss rates(1)

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

amm

p_2k

appl

u_2k

art_

2k

bzip

2_2k

craf

ty_2

k

eon_

2k

equa

ke_2

k

fma3

d_2k

galg

el_2

k

gcc2

k

gzip

_2k

luca

s_2k

mcf

_2k

mes

a_2k

mgr

id_2

k

pars

er_2

k

swim

_2k

twol

f_2k

vorte

x_2k

vpr_

2k

wup

wis

e_2k

Programs

Cach

e m

iss-

rate

MeasuredPredicted

Method 1: Using weights for prediction

Note: Assume every program to be a customer application one at a time

24

Prediction of data cache miss rates(2)

Method 2: Using clustering for prediction

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

amm

p

appl

u

art

bzip

2

craf

ty

eon

equa

ke

fma3

d

galg

el

gcc

gzip

luca

s

mcf

mes

a

mgr

id

pars

er

swim

twol

f

vorte

x

vpr

wup

wis

e

Programs

Cach

e m

iss-

rate

Measured

Predicted

Note: Assume every program to be a customer application one at a time

25

Outline Motivation and Objectives

Methodology

Experimental results

Conclusion

Future work

26

Conclusion Demonstrated two simple methods to predict

performance

Used SPEC CPU2000 as an example to predict performance.

The accuracy of prediction depends on two factors:

How well the workload characteristics correlate to performance Is there a program similar to the customer application in the

repository of known programs

27

Future Work

Two main items on the TO DO list:

To add more programs to the repository and validate the results

To calibrate the measure of similarity (distance) in workload space to the error in the target metric space.

28

Thank you !!