Liquid Architecture

D. Schuehler, B. Brodie, R. Chamberlain, R. Cytron,

S. Friedman, J. Fritts, P. Jones, P. Krishnamurthy,

J. Lockwood, S. Padmanabhan, and H. Zhang

Dept. of Computer Science and EngineeringWashington University in St. Louis

Supported by NSF ITR-0313203

Microarchitecture Optimizationfor Embedded Systems

Liquid Architecture

• Configurable architecture that can adapt to needs of particular application

• E.g., within an FPGA– Soft-core processors

• E.g., as an embedded processor– Tensilica supports configuration at fab

time– Stretch support configuration at run time

• Today’s discussion is on performance analysis and configuration choice

Block Diagram

`̀`̀Layered Internet Protocol Wrappers

ControlPacket

Processor

ExternalMemory

AHB

APB

I-Cache D-Cache

UARTUARTLEDLED

AdapterAdapter BootRom

FPGA

StatisticsModule

NetworkInterface

FPXEvent Bus

MemoryController

LEONSPARC-

compatibleprocessor

Microarchitecture Configurability

• Instruction set• Memory subsystem

– Cache size (I and D)– Associativity– Cache line size

• Co-processor(s)• Instruction pipeline• Full HDL source is available

Design Flow

Write and compile embedded SPARC application with GCC

Internet

Reconfigure FPX hardware via Internet and upload system software.

Identify configuration for candidatearchitecture

Execute program on FPX Platform and measure run-time performance

Cycle-accurate profiling

MethodTime / Cycles

.text

main

addQuery

findMatch

computeKey

computeBase

computeStep

fillQuery

Rnd

• Choose methods to profile from the user interface

MethodAddressRange

.text

main

addQuery

findMatch

computeKey

computeBase

computeStep

fillQuery

Rnd

0x400003EFHi

0x4000027CLo

Method

.text

main

addQuery

findMatch

computeKey

computeBase

computeStep

fillQuery

Rnd

0x400003EFHi

0x4000027CLo

0x4000035A

Statistics Module

PC CLK

Event Bus

Function

.text

main

addQuery

findMatch

computeKey

computeBase

computeStep

fillQuery

Rnd

0x400003EFHi

0x4000027CLo

0x4000035A≤ ≤

Counter

Statistics Module

PC CLK

Event Bus

INCR

Function

.text

main

addQuery

findMatch

computeKey

computeBase

computeStep

fillQuery

Rnd

0x400003EFHi

0x4000027CLo

0x4000035A≤ ≤

Counter

PC CLK

0x4000061FHi

0x400005D8Lo

0x4000035A≤ ≤

Counter

Statistics Module

Event Bus

INCR

INCR

0x400003EFHi

0x4000027CLo

0x4000035A≤ ≤

Counter

PC CLK

0x4000061FHi

0x400005D8Lo

0x4000035A≤ ≤

Counter

Statistics Module

Event Bus

To User

INCR

INCR

Where is time spent?

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

128K 32K

Size of hash table (Bytes)

% o

f to

tal ru

nti

me Rest

coreLoopfindMatch BLASTN

biosequencesearchapplication

FunctionTime / Cycles

.text

main

addQuery

findMatch

computeKey

computeBase

computeStep

fillQuery

Rnd

Cache Hits / Misses

Read Write

Expand tomeasure cachehits/misses

Measure Several Configurations

Impact of D-cache Configuration

BLASTNbiosequencesearchapplication

86

88

90

92

94

96

98

100

128K, 1Kx1 128K,32Kx1

128K,16Kx2

32K, 1Kx1 32K, 32Kx1 32K, 16Kx2

Size of hash table, D-cache configuration

hit

rat

e (%

)

TotalfindMatchcoreLoop

Impact of I-cache Configuration

BLASTNbiosequencesearchapplication

0

5

10

15

20

25

30

35

128K 32K

BLASTN hash table sizes (Bytes)

Ru

n t

ime (

secs)

1KB I -Cache4KB I -Cache

FunctionTime / Cycles

Cache Hits / Misses

Read Write

.text

main

addQuery

findMatch

computeKey

computeBase

computeStep

fillQuery

Rnd

PipelineStalls

BranchPredict

Time for Single Run

Almost 2orders ofmagnitudefaster thansimulation

80000

1800

1

10

100

1000

10000

100000

SimpleScalar 3.0 LEON

Tim

e (

sec)

Implications of Slow Simulation

• Focus has historically been on measuring the performance of a single thread of a single application

• Real apps are often executed in a multitasking environment– Impacts cache behavior– Ignores OS (system call) performance

• Liquid architecture system enables direct measurement, including OS

OS Boot Sequence

Summary

• Run-time reconfigurable processors will be available sooner rather than later

• Determining desired configuration is a difficult design task– Large search space– Depends on accurate performance data

• Liquid architecture system enables direct measurement of performance properties

Current and Future Work

• Evaluation of several arch. design ideas• Automated search of the design space• Characterizing performance analysis

methods– Analytic models– Simulation models– Direct execution models

• Usable as is for evaluating soft-core procs

• Like to extend to higher-speed procs