Cyrus: Unintrusive Application-Level Record-Replay for Replay Parallelism

Transcription

1 Cyrus: Unintrusive Application-Level Record-Replay for Replay Parallelism Nima Honarmand, Nathan Dautenhahn, Josep Torrellas and Samuel T. King (UIUC) Gilles Pokam and Cristiano Pereira (Intel) iacoma.cs.uiuc.edu 1

2 Record-and-Replay (RnR) Record execution of a parallel program or a whole machine Save non-deterministic events in a log During replay, use the recoded log to enforce the same execution Each thread follows the same sequence of instructions Use cases Debugging Security High availability 2

3 Contribution: Cyrus RnR System Application-level RnR RnR one or more programs in isolation What users typically need Fast replay Replay-time parallelism Flexibly trade off parallelism for log size Unintrusive HW No changes to snoopy cache coherence protocol 3

4 Capturing Non-determinism Sources of non-determinism Program inputs Memory access interleavings How to capture? OS kernel extension to capture program inputs HW support to capture memory interleavings (HW-assisted RnR) This talk: recording memory interleavings 4

5 Time Recording Interleaving as Chunks Inter-processor data dependences manifest as coherence messages Capture interleavings as ordered chunks of instructions P0 add. store A. mul. sub Req Resp P1 div. load A. add. P0 add... store A... mul. sub P1 div. load A. add 5

6 Restriction: Unintrusive HW Unmodified snoopy protocols In some coherence transactions, there is no reply P0 P1 P0 invl P1 P0 invl P1 data P1 rd data P1 wr P1 wr RAW WAW Requirements for HW-assisted RnR: Do not augment or add coherence messages Do not rely on explicit replies Only source is always aware Use source-only recording 6 WAR

7 Challenge 1: Enable Replay Parallelism Key to fast replay Overlapped replay of chunks from diff. threads Previous work: DAG-based ordering (Karma [ICS 2011]) Requires explicit replies Augments coherence messages 7

8 Challenge 1: Enable Replay Parallelism P0 P1 P0 P1 P0 P1 P0 P1? P0 P1 Predecessor Successor 8

9 Challenge 2: Application-Level RnR Turn hardware on only when a recorded application runs. P0 P1 P2 Four cases: (1) src=monitoring, dst=monitoring (2) src=monitoring, dst=not monitoring (3) src=not monitoring, dst=monitoring (4) src=not monitoring, dst=not monitoring (1) (2) (3) (4) Issues of source-only recording: Cannot distinguish between (1) and (2) (2) may result in a dependence later Not recording in (3) and (4) Non-monitored Communication Monitored Application 9

10 Challenge 2: Application-Level RnR Treat (2) as an Early Dependence Defer and assign it to the next chunk of the target processor (3) and (4) superseded by context switches At context switch, record a Serialization Dependence to all other processors P0 P1 P2 (1) ser ser (3) (2) (4) Non-monitored Monitored Application Dependence 10

11 Key: On-the-Fly Backend Software Pass Recording Processors P P P On-the-fly Backend Replaying Processors P P P Sourceonly Log DAG of Chunks Transforms source-only log to DAG (for parallelism) Fixes the Early and Serialization dependences To support app-level RnR Can trade replay parallelism for log size 11

12 Memory Race Recording Unit (RRU) HW module that observes coherence transactions and cache evictions Tracks loads/stores of the chunk in a signature Keeps signatures for multiple recent chunks Records for each chunk # of instructions Timestamp (# of coh. transactions) Dependences for which the chunk is source Dumps recorded chunks into a log in memory P Cache Bus Mem Refs Evictions Snoops RRU 12

13 TimeStamp RRUs Record Source-Only Log C00 C01 P0 P1 P2 Rd A Wr B C10 Wr A Rd D C20 Wr D Rd B P0 P0 P1 P2 C C P1 C P2 Chunk TS Successor Vector C

14 Backend Pass Creates DAG Finds the target chunk for each recorded dependency Creates bidirectional links between src and dst chunks This algorithm is called MaxPar C00 Chunks of P0 Chunks of P1 C C C C01 C10 Chunks of P2 C C20 14

15 Trading Replay Parallelism for Log Size C00 C20 C01 C10 C20 C00 + C01 C10 CPU TID SIZE PTV STV TID SIZE C00 + C01 C10 CPU TID SIZE PTV STV C C00 No Parallelism Smallest log Less Parallelism Smaller log TID SIZE C01 C10 No Parallelism Even Smaller log 15 C20

16 Evaluation Using Simics Full-system simulation with OS Wrote a Linux kernel module to Records application inputs Controls RRUs Model processors 8 processors for the app 1 processor for the backend 10 SPLASH-2 benchmarks 16

17 Normalized Replay Time (Log Scale) Replay Time Normalized to Recording rep-maxpar rep-stitched rep-stserial rep-serial Large difference between MaxPar and Serial replay On 8 processors, unoptimized MaxPar replay is only 50% slower than recording 17

18 Conclusions Cyrus: RnR system that supports Application-level RnR Unintrusive hardware Flexible replay parallelism Key idea: On-the-fly software backend pass On 8 processors: Large difference between MaxPar and Serial replay Unoptimized replay of MaxPar is only 50% slower than recording Negligible recording overhead Upcoming ISCA 13 paper describes our FPGA RnR prototype 18