cisc 879 - machine learning for solving systems problems arch explorer lecture 5 john cavazos dept...

CISC 879 - Machine Learning for Solving Systems Problems

Arch ExplorerLecture 5

John CavazosDept of Computer & Information Sciences

University of Delaware

www.cis.udel.edu/~cavazos/cisc879


Motivation

2

0,98

0,99

1,00

1,01

1,02

1,03

1,04

1,05

1,06

1,07

1,08

1,09

1,10

1,11

VC('82)

TP('8

2)

SP('9

0)

Mar

kov(

'97)

FVC(

'00)

DBCP('0

1)

TKVC('0

2)

TK('0

2)

CDP('0

2)

CDPSP

('02)

TCP('0

3)

GHB('0

4)

Mechanism

Avera

ge S

peed

up

[MICRO 2004, Gracia-Pérez et al.]

Need for systematic quantitative comparison

Need for systematic quantitative comparison


Computer Arch Research

3


Design space exploration

4

Time-to-marketTime-to-marketarea

power

execution time

Multi-objectivesMulti-objectives

Need more than intuition and experience!


ArchExplorer

5

archexplorer.org database

simulation cluster

upload

daily update

pick design points

add results

test

Server-side InfrastructureWebsite

FULLY AUTOMATIC


How to compare?

1. Custom simulator2. Hardware compatibility3. Software compatibility4. Upload

6

Wrapped Simulator& Parameter ranges

CustomSimulator

DL1

CPU

DSM

EX WB CMM

F

$ TLB

$

$ TLB

$

MEM

D S EX WB CMM

FSM

IL1BP

L2

S

MEM


Hardware compatibility

7

Instruction caches Data cachesBranch predictorsInterconnectsMain memoryAccelerators...


Software compatibility

8

Isolate the hardware block, possibly by from centralized control to distributed control


Software compatibility

9

Self-Configuration and parameters legality

Models of computationWrapping in SystemC-based on UNISIM communication layer


Case study

Memory sub-system for embedded processor• PowerPC405• 8 different cache

modules available• Complex hierarchies

automatically explored

• Ranking designs for performance, power, energy, area,...

Victim Cache

Timekeeping Victim cache

Stride Prefetcher

Content-Directed Prefetcher

Stride + Content Directed Prefetcher

Tag Prefetcher

Global History Prefetcher

Skewed associtiative cache

10


Accurate comparison needs compiler tuning as well

11

1.231.23

2.622.62

1.091.09

P1P1 P2P2<

P1P1 P2P2>

baseline

Tuned to P1, tuned to P2


Best data cache mechanisms per area

12

CONCLUSIONS:1. Contrast to

Gracia-Pérez et al. [MICRO 2004]

2. No clear winner3. Close to tuned

parametric cache


Best data cache mechanisms per area

13

CONCLUSIONS:1. Contrast to

Gracia-Pérez et al. [MICRO 2004]

2. No clear winner3. Close to tuned

parametric cache


Composing cache hierarchies

14


Speedup and Energy Improvement

15


ARCHEXPLORER.ORGCheck out this website:

16

CISC 879 - Machine Learning for Solving Systems Problems 17


Conclusion

• Permanent open competition(s)• Future:

• superscalar processor• branch predictor repository• multi-cores

• Open for your ideas!• NoC, compiler extensions,...

18


ARCHEXPLORER.ORGCheck out this website:

19


Genetic Search Algorithm

Convergence

• Permanently ranks all designs • per area bucket• speedup or power• assigning higher

probability to better points

• Picking a point according to distribution

• Mutations & crossover• Natural selection

20Veerle Desmet – Sylvain Girbal – Olivier Temam6th HiPEAC Industrial Workshop – Thales Nov 26th, 2008

StatisticalExploration

$

BPBP

CPU$$

MEM


Standardized Interfaces

ModuleRepository

Features for Systematic DSE

Module parameter tuningModule explorationCompiler

Exploration

Design Space Exploration

CompatibilityDatabase

ParameterCheck

ParameterIntrospection

Compatibilitydatabase

Compiler FlagDatabase

benchmarksdatasets

PPC ARM

WB$ VC$ SP$NBWB$ TVC$

CDP$ CDPSP$ TagP$ GHB$

BUS

DRAM

Module categoryModule interfaces

Known models

Probing neighborsparameters

Configuration validityRanges

Params. relationship

DRAMnBanks {2;4;8}tRAS+tCD<tRCD

focused search algorithm

configsSelection probabilityFast convergence

configsconfigs

Predictive modeling

compilerflags

Machine description

http://unisim.org

cisc 879 - machine learning for solving systems problems arch explorer lecture 5 john cavazos dept...

Documents