NC STATE UNIVERSITY ASPLOS-XII Understanding Prediction-Based Partial Redundant Threading for Low-Overhead, High-Coverage Fault Tolerance Vimal Reddy Sailashri.

NC STATE UNIVERSITYASPLOS-XII

Understanding Prediction-Based Partial Redundant Threading for Low-Overhead, High-Coverage

Fault Tolerance

Vimal ReddySailashri Parthasarathy†

Eric Rotenberg

Dept. of Electrical & Computer EngineeringNorth Carolina State UniversityRaleigh, NC

† Architecture Modeling Infrastructure Group Intel Corporation Hudson, MA

© 2006 Vimal Reddy

2


Transient Fault Tolerance

• Transient faults– Temporary hardware faults– Worsening with shrinking technology

• Soft errors• Noise

• Prominent solution: Redundant Multithreading– Full program duplication– Complete fault tolerance– High overheads

© 2006 Vimal Reddy

3


Full Redundant Execution

• Full duplication• 100% fault coverageA(f1)

B(f1)

Main Thread

B(f2)

A(f2)

Redundant Thread

!=

fault

fault

!=

detected

detected

© 2006 Vimal Reddy

4


Partial Redundant Threading (PRT)

• Only partially duplicate a program• Shorter the redundant thread, lesser the

overhead• Approach taken to create partial thread affects

fault tolerance

© 2006 Vimal Reddy

5


fault

!=

Conventional PRT

Partial Redundant

Thread

copy state

detected

A(f)

B(f)

Main Thread

B(p)

© 2006 Vimal Reddy

6


fault

Conventional PRT

B(p)

copy state

==

• Arbitrary duplication• Non-duplicated portions

lose fault coverage

undetected

A(f)

B(f)

Main Thread

Partial Redundant

Thread

© 2006 Vimal Reddy

7


Prediction-based PRT

copy state

PAA(f)

B(f)

Main Thread

Confident prediction of A

Freq. Case: Correct prediction(e.g., 99.9% of the time)

Partial Redundant

Thread

B(p)

fault

!=detected

© 2006 Vimal Reddy

8


fault


PAA(f)

B(f)

Main Thread

!=detected

Partial Redundant

Thread

B(p)

Confident prediction of A

Freq. Case: Correct prediction(e.g., 99.9% of the time)

© 2006 Vimal Reddy

9


fault


PAConfident prediction of A

Rare case: Incorrect prediction(e.g., 0.1% of the time)

A(f)

B(f)

Main Thread

==

undetected

Partial Redundant

Thread

B(p)

© 2006 Vimal Reddy

10



PAA(f)

B(f)

Main Thread

• Confident predictions are good proxies for redundant execution

• Predictions break thread inter-dependence

• Near-100% fault coverage

Partial Redundant

Thread

B(p)

© 2006 Vimal Reddy

11


Relaxing checking constraints

Partial Redundant Thread

Main Thread

- S(p) and T(p) are removableS(f)

T(f)

A(f)

B(f)

PA

B(p)

S(p)

T(p)

© 2006 Vimal Reddy

12


Relaxing checking constraints

Partial Redundant Thread

Main Thread

- S(p) and T(p) are removable

- Such removal can shorten partial thread significantly

S(f)

T(f)

A(f)

B(f)

PA

B(p)

S(p)

T(p)

How to check S(f) and T(f)?

Checks both S(f) and T(f)

- But, no predictions to replace them

fault

!=detected

fault

© 2006 Vimal Reddy

13


PRT SpectrumPARTIAL REDUNDANT THREADING (PRT) SPECTRUM

ConfidentPredictions

PartialDuplication

PartialDuplication

&Confident

Predictions

EX: Opportunistic EX: Slipstream EX: ReStore

• Case study: PRT on Slipstream

M. Gomaa T. N. Vijaykumar

ISCA 2005

Z. PurserK. Sundermoorthy E. Rotenberg

MICRO 2000 ASPLOS 2000

N. Wang S. J. Patel

DSN 2004

© 2006 Vimal Reddy

14


Slipstream Overview

DelayBuffer

Branch + Valueoutcomes

verifySMT

Processor

VerifiedArchitectural

State

R-stream

UnverifiedArchitectural

State

A-stream

Outcome mismatch

copy

restart

© 2006 Vimal Reddy

15


Slipstream Overview

DelayBuffer


verify


State

R-streamA-stream


State

Slipstream Components

remove

• Detect predictable instructions

• Predict based on repeated detection (confidence)

SMT Processor

monitor

© 2006 Vimal Reddy

16


Slipstream Overview

DelayBuffer


verify


State

R-streamA-stream


State


remove

SMT Processor

Predictions act as proxies



monitor

© 2006 Vimal Reddy

17


Slipstream Overview

DelayBuffer


verify


State

R-streamA-stream


State


remove

• Confident branches

• Confident dead writes

• Confident silent writes

• Slices whose leaves are any of the above

SMT Processor

Predictions act as proxies



monitor

© 2006 Vimal Reddy

18


Confident Branch

detected

!=Correct Branch Prediction (Taken)

Main (R-stream) Partial (A-stream)

(Not Taken)

H(f)

Branch

fault

H(p)

Branch

© 2006 Vimal Reddy

19


Confident BranchMain (R-stream) Partial (A-stream)

(Not Taken)

H(f)

Branch

fault

Incorrect Branch Prediction (Not Taken)

==

undetected

© 2006 Vimal Reddy

20



fault

DeadWrite

H(f)

not detected, but safe

Correct Prediction(Dead)

DeadWrite

H(p)

Confident Dead Write

© 2006 Vimal Reddy

21



fault

DeadWrite

H(f)

Incorrect Prediction(Not Dead)

J(f)

?

==undetected

Confident Dead Write

J(p)

© 2006 Vimal Reddy

22



Confident Silent Write/Store

SilentWrite

H(f)

72

SilentWrite

H(p)

72

J(p)J(f)

7272

© 2006 Vimal Reddy

23




SilentWrite

H(f)

3

SilentWrite

H(p)

3

J(p)J(f)

33

© 2006 Vimal Reddy

24




SilentWrite

H(f)

V

SilentWrite

H(p)

J(p)J(f)

VV

Correct Prediction(Silent)

fault

XVX

detected!=

© 2006 Vimal Reddy

25




SilentWrite

H(f)

stale

J(p)J(f)

Incorrect Prediction (Not Silent)

fault

X

== undetected

© 2006 Vimal Reddy

26


Slipstream Fault Detection Coverage

• We showed confident predictions detect faults– Additional coverage =

Correctly predicted confident instr. + Their backward slices– Mispredictions vulnerable, but rare in Slipstream– Mispredictions + backward slices = only 0.1% instr.

• Hence, non-duplicated instr. well covered

Slipstream has high fault detection coverage (99.9%)

• Prior research: Only duplicated instr. covered

© 2006 Vimal Reddy

27


Fault Coverage: Detection vs. Recovery

Detection Coverage

• Represents ability to detect faults

• Slipstream: 99.9% of instr.

Recovery Coverage

• Ability to rollback to ‘golden’ state on fault detection

• Slipstream’s recovery coverage?

© 2006 Vimal Reddy

28


Slipstream Fault RecoveryPartial Redundant

ThreadMain

Threadfault

detectedFlawedRollback(Conventional Slipstream)

CorrectRollback

!=A(f)

B(f)

C(f)

S(f)

T(f)

PA

B(p)

C(p)

© 2006 Vimal Reddy

29


Reorder Buffer (ROB)

faulthead

commit to arch. state

tail

S(f)

T(f)

A(f)

B(f)

C(f)

PAA(f) !=

FlushRestart

Arch. State Corrupted

Base SlipstreamRecovery

detected

© 2006 Vimal Reddy

30



headcommit to arch. state

tail

S(f)

T(f)

A(f)

B(f)

C(f)

PAA(f) !=

Flush to ROB headRestart

fault

Enhancement:ROB-head Recovery

detected

© 2006 Vimal Reddy

31




tail

S(f)

T(f)

A(f)

B(f)

C(f)

PAA(f) !=

Flush to ROB head

fault

Enhancement:ROB-head Recovery

detected

Arch. State Corrupted

Limited rollback distance: • R-stream retires quickly – accelerated by leading A-stream

© 2006 Vimal Reddy

32




tail

S(f)

T(f)

A(f)

B(f)

C(f)

PAA(f) !=

Enhancement:ROB occupancy management

R(f)

P(f)

Threshold

• Delay in retirement, hence performance hit

Increased rollback distance

detected

© 2006 Vimal Reddy

33




tail

S(f)

T(f)

A(f)

B(f)

C(f)

PAA(f) !=

fault

Enhancement:History Buffer

History Buffer

headtail

Undo changes

Performance friendly: R-stream mispredictions are rare

© 2006 Vimal Reddy

34



Indirect Check of Silent Write/Store

SilentWrite

H(f)

V

J(p)J(f)


fault

detected !=

X


© 2006 Vimal Reddy

35



Direct Check of Silent Write/Store

SilentWrite

H(f)

V

J(p)J(f)


fault

X

existing value

new value!= V

Direct Check


detected

Need Enhanced Recovery

© 2006 Vimal Reddy

36


Novel Framework to Analyze Coverage

• Each instr. considered “candidate faulty”

• Coverage = # of instr. checked before committal to arch. state

• Mispredicted instr. and backward slices marked unchecked

© 2006 Vimal Reddy

37


Prior Work New Analysis

Checkers

Non-checkersCoverageAnalysis

A

R

Dup

A

R

Dup

A

R

Dup

R

Conf. Dead

A pred.

R

Conf. Br.

R

Conf. SW

R

NonDup

R

NonDup

.R

Conf. SW

A pred

© 2006 Vimal Reddy

38


Logically masked Masked?

CoverageAnalysis

Rollback point

A

R

Dup

A

R

Dup

A

R

Dup

R

Conf. Dead

A pred.

R

Conf. Br.

R

Conf. SW

R

NonDup

R

NonDup

Masked?

.R

Conf. SW

A pred

© 2006 Vimal Reddy

39


Clarification

• Analysis framework is a coverage measurement tool

• Not in the actual hardware

© 2006 Vimal Reddy

40


Results: Microarchitecture models

L1 I & D caches

64KB, 4-way, 64B line, LRU,L1hit = 1 cycle, L1miss/L2hit = 10 cycles

L2 unified cache

1MB, 8-way, 64B line, LRU,L1miss/L2miss = 100 cycles

superscalar core

dispatch/issue/retire bandwidth: 8 (4)reorder buffer (ROB): 256 (128)load/store queue: 64 (32)issue queue: 64 (32)cache ports (read/write): 4 (2)

© 2006 Vimal Reddy

41


Breakdown of Instructions

SPEC2K Integer Benchmarks

0

10

20

30

40

50

60

70

80

90

100

bzip gap gcc gzip mcf parser perl twolf vortex vpr

% o

f to

tal i

nst

ruct

ion

s

Non-Dup

Dup

AVG

65% duplicated

35% non-duplicated

> 50% non-duplicated on some bench.

© 2006 Vimal Reddy

42


Breakdown of Instructions

0

10

20

30

40

50

60

70

80

90

100

bzip gap gcc gzip mcf parser perl twolf vortex vpr

% o

f to

tal i

nst

ruct

ion

s

bs_Conf_Pred

Conf_Pred

Dup

AVG

23% confident predictions

12% backward slices of confident predictions

See paper for more detailed breakdown

© 2006 Vimal Reddy

43


Fault Coverage

0

10

20

30

40

50

60

70

80

90

100

PriorWork

Slip Slip +Direct

RH0 RH32 RH48 HB16 HB32

% f

ault

co

vera

ge

otherbs_SWbs_SSbs_BSWSSDBDup

Slip + Direct +Recovery Enhancements

Prior Work (65%): Only dup. instr.

New result (78%): Correct confidentpredictions covered

ROB-head recovery improves with occupancy threshold: 95% to 98%

Direct silent write/ store checks improve coverage(88%)

65%

78%

88%

95% 97% 98% 98% 99%

History buffer schemes provide high coverage (98% to 99%)

© 2006 Vimal Reddy

44


Slipstream Performance (SMT 8-wide)

Average slowdown: Full redundant execution:14%

Slipstream:1.3%0

1

2

3

4

bzip

gap

gcc

gzip

mcf

pars

er perl

twolf

vorte

xvp

rAvg

IPC

Single

Full

Slip

© 2006 Vimal Reddy

45


Performance Impact of Enhanced Recovery

ROB-occupancy management delays retirement, causes slowdown(gradual decrease from RH0 to RH48)

History buffer approach is performance friendly (negligible slowdown)

0

1

2

3

4

bzip

gap

gcc

gzip

mcf

pars

er perl

twolf

vorte

xvp

rAvg

IPC

SlipSlip:RH0Slip:RH32Slip:RH48Slip:HB16Slip:HB32

© 2006 Vimal Reddy

46


Conclusions

• Confident predictions can replace duplication– Slipstream case study : Redundant thread

reduced by up to 57% while retaining near-100% coverage

• Prediction-based PRT offers a new avenue for efficient fault tolerant computing

NC STATE UNIVERSITY ASPLOS-XII Understanding Prediction-Based Partial Redundant Threading for Low-Overhead, High-Coverage Fault Tolerance Vimal Reddy Sailashri.

Documents

nc state university

fault coverage fault

fault tolerance slide

time fault slide

fault conventional prt

time slide

removable slide

ma slide