SELSE ASAR: Applica+on-Specific Approximate Recovery to Mi+gate Hardware Variability. Presenter: Manish Gupta

Transcription

1 SELSE 2017 ASAR: Applica+on-Specific Approximate Recovery to Mi+gate Hardware Variability Presenter: Manish Gupta Collaborators: Abbas Rahimi, Daniel Lowell, John Kalama9anos, Advisors: Dean Tullsen, Rajesh Gupta 1

2 Variability Greater Performance Lower Power Higher Varia+ons Process Aging Temperature Vcc Droops 2

3 Mi+ga+ng Variability Effects Detect Errors in Hardware Ignore Fix 3

4 Today Detect Errors in Hardware Ignore Fix Today, we detect and fix in hardware. Using hardware error recovery. 4

5 This Work Detect Errors in Hardware Ignore (Discard) SoUware Fix (Rerun) In this work, we champion sopware recovery [1]. [1] Relax: An Architectural Framework for SoUware Recovery of Hardware Faults, ISCA

6 Ignore (Discard) vs. Fix (Rerun) Detect Errors in Hardware Ignore (Discard) SoUware Fix (Rerun) Discard may lead to unacceptable output quality. Rerun suffers from high recovery overhead. 6

7 Outline Variability Problem Hardware/SoUware Recovery SoUware Recovery (Rerun or Discard) [Relax ISCA 10] Emerging Approximate Applica9ons ASAR: Applica9on Specific Approximate Recovery Hybrid Recovery (ASAR + Rerun) Summary 7

8 Emerging Approximate Applica+ons Data-intensive Scien9fic-Compu9ng Vision & media 8

9 Approximate Recovery Detect Errors in Hardware Ignore (Discard) Approximate Recovery SoUware Fix (Rerun) ASAR: Applica9on-Specific Approximate Recovery - Worst output quality - Best recovery performance - Good output quality - Good recovery performance - Best output quality - Worst recovery performance 9

10 HW/SW Organiza+on Main contribu9ons of this work is SW-based recovery using: 1) Rerun 3) ASAR (Approx. recovery) 2) Discard 4) Hybrid (Rerun + ASAR) SoUware Hardware Cri9cal Try { Non-cri9cal } Non-cri9cal Catch { SW recovery } HW error detec9on [2, 3] Safe mode Unsafe mode [1] Expose HW-error info. [1] Truffles: Architecture support for disciplined approximate programming. H. Esmailzadeh, A. Sampson, L. Ceze, D. Burger. In ASPLOS 12 [2] Argus: Low-cost comprehensive error detec9on in simple cores. A. Meixner, M. E. Bauer, and D. J. Sorin. In MICRO 2008 [3] Relax: An Architectural Framework for SoUware Recovery of Hardware Faults M. Kruijf, S. Nomura, K. Sankralingam. In ISCA 10 10

11 ASAR: Applica+on-Specific Approximate Recovery ASAR Interpolate 1 ASAR Sampling 2 ASAR Reuse 3 1. P. Whatmough, S. Das, and D. Bull. Hybrid circuit and algorithmic 9ming error correc9on for low-power robust dsp accelerators A-SSCC, S. Agarwal BlinkDB: Queries with Bounded Errors and Bounded Response Times on Very Large Data EuroSys Y. Ding et al., A Compiler Scheme for Reusing Intermediate Computa9on Result CGO,

12 Try-Catch Refactoring: ASAR Reuse K-means for N points [P1, P2,,PN]. Assign the closest of M center [C1, C2,.., CM]. Cri9cal Try { Try { Non-cri9cal } } Catch Catch { ASAR { ASAR Reuse} Reuse} Next, two ques9ons we explore for all applica9on are: 1) The dynamic ra9o of Cri9cal vs. Noncri9cal code. 2) Absolute performance gain of using approximate recovery over rerun. In each itera+on For Pi in [P1,..,PN]{ Try { dmin = INT_MAX; for c in [C1, C2,.., CM]: d = distance(pi, c); if (d < dmin) { Cen[Pi] = c; ReuseBuf[Pi] = c; } } Catch { Cen[Pi] = ReuseBuff[Pi] } } //for pi in [P1,..,PN] 12

13 Answering 1) and 2) 1) Cri9cal vs. Non-cri9cal 2) Catch { ASAR } Catch { Rerun } 13

14 Evalua+on Six applica9on from various domains. SW-based using 1) rerun, 2) discard, 3) ASAR, and 4) hybrid (ASAR + Rerun) recovery. Random fault-injec9on simulated. Finally, we measure cpu cycles spent in cri9cal, non-cri9cal, and recovery code regions. 14

15 ASAR vs. Rerun 15

16 K-Means 16

17 K-Means 17

18 K-Means 18

19 ASAR vs. Rerun 19

20 Hybrid (ASAR + Rerun) Recovery ASAR only provides one addi9onal opera9ng point. Only approximate recovery could degrade below user acceptable level. Hybrid recovery to explore performance-quality at a finergranularity and allow user to specify desired output quality. Performance Training Input [fi, rj] Check (QoS obs > QoS thd) Discard ASAR Rerun Output quality (QoS) QoS Model Genera*on Try { non-cri)cal} Catch{ 1. ASAR 2. Rerun } Hybrid Recovery QoS Model U*liza*on QoS mod (fi,rj):qos QoS thd Test Input 20

21 Hybrid (ASAR + Rerun) Recovery 21

22 K-means 22

23 K-means 23

24 K-means 24

25 Hybrid (ASAR + Rerun) Recovery 25

26 Summary We propose ASAR, applica9on-specific approximate recovery. To guarantee the output acceptability and explore performance-quality curve at a finer-granularity we also propose a hybrid recovery mechanism. We show that, depending on the applica9on characteris9cs, approximate recovery can reduce recovery overhead and improve output quality rela9ve to rerun and discard, respec9vely. 26

27 Thanks! 27