1 / 75

Hardware Memory Race Recording for Deterministic Replay

Hardware Memory Race Recording for Deterministic Replay. Mark D. Hill University of Wisconsin—Madison August 10, 2007 Based on joint work with Min Xu & Ras Bodik: ISCA 2003, ASPLOS 2006, IEEE Micro Top Picks 2007, & Xu UW Ph.D. 5/2006 (slides updated from defense talk).

manueli
Download Presentation

Hardware Memory Race Recording for Deterministic Replay

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Hardware Memory Race Recording for Deterministic Replay Mark D. Hill University of Wisconsin—Madison August 10, 2007 Based on joint work with Min Xu & Ras Bodik: ISCA 2003, ASPLOS 2006, IEEE Micro Top Picks 2007,& Xu UW Ph.D. 5/2006 (slides updated from defense talk).

  2. Wisconsin Multifacet Project • Seek improved architectures for (mostly) servers thatare (mostly) chip multiprocessors (CMPs, multi-core) • Led by Mark Hill & David Wood • LogTM work w/ Ben Liblit & Mike Swift • Funding • Grants from U.S. National Science Foundation • Donations from Intel and Sun

  3. Selected Multifacet Results (1 of 2) • Multiprocessor Flight Data Recorder • Records memory races for deterministic replay • Piggyback on coherence protocol & logs 0.001B/instrn • Supports SC & TSO • Adaptive L2 Cache & Memory Link Compression • Cache compression creates level 2½ cache (or 3½) • Adaptive so as “to do no harm” • Link compression husbands memory link bandwidth • Multifacet GEMS MP Simulation Infrastructure • Simics==Correctness; GEMS==Performance • GPL Distribution

  4. Selected Multifacet Results (2 of 2) • Log-based Transactional Memory (LogTM) • Accelerates commit by writing new values in place(after saving old values in a per-thread log) • Gracefully handles cache eviction of TM data • LogTM Signature Edition (LogTM-SE) • Signatures summarize read/write sets • HW mechanisms: simple, policy-free, SW accessible • Forthcoming • Mechanisms to handle thread switching/migration & paging of transactions with OS or OS/VMM

  5. Effective Inexpensive Long Recording More Applicable Low Overhead Low Cost Race Recorder Overview • Increasingly useful to replaymultithreaded code • Race recording: key to dealing with nondeterminism • A Case Study • Long recording: 1 byte/kilo-instr • Always-on recording: less than 2% overhead • Low cost: 24 KB RAM/core • Support both SC & TSO (x86-like)

  6. Contributions Low Runtime Overhead Small Log Size Coherence Piggyback Transitive Reduction & Regulated TR Effective Inexpensive Order-Value Hybrid Set/LRU Approximation Low Cost Hardware SC & TSO Applicability

  7. Outline 6 slides Motivation & Problem 21 An Effective and Inexpensive Race Recorder TR & RTR Algorithms Coherence Piggyback Set/LRU Approximation Order-Value Hybrid 6 Evaluation Method & Results 3 Conclusions, etc.

  8. Motivation & Problem

  9. Multithreaded Debugging • % gdb a.out • gdb> run • Program received SIGSEGV. • In get() at hash.c:45 • 45 a = bucket->d; • % gcc hash.c • % a.out • Segmentation fault • % • % gcc para-hash.c • % a.out • Segmentation fault • % • % gdb a.out • gdb> run • Program exited normally. • gdb> • % gcc para-hash.c • % a.out • Segmentation fault • Race recorded in “log” • % • % gdb a.out log • gdb> run • Program received SIGSEGV. • In get() at para-hash.c:67 • 67 a = bucket->d;

  10. Applications of Deterministic Replay • Deterministic Replay is logically recreating a program execution • Cyclic Debugging ([Pancake & Netzer ‘93]) • Fault Tolerance (ExtraVirt [Lucchetti et al. ’05]) • Intrusion Analysis (ReVirt [Dunlap et al. ’02]) • Data Recovery (Continuous Checkpointing)? • See VMware Workstation 6 • Replay included for single-processor guest VM

  11. Log - X = X*5 - - Recording X= 6 Race Recording Thread I Thread J Thread I Thread J X = 1 X++ print(X) - - - X = X*5 - - X = X*5 - - X = 1 X++ print(X) Original Replay X=6 X=10

  12. Focus Recording for Multithreaded Replay • Race Recording • Not-an-issue for a single thread • Create the same general & data races • Checkpointing • Provide a snapshot of the program state • Many proposals (e.g., SafetyNet), not focus • Input Recording • Provide repeatable inputs • Some proposals (e.g., part of FDR), not focus

  13. A Good Race Recorder Low runtime overhead Applicability Low cost • % gcc para-hash.c • % a.out • Segmentation fault • Race recorded in “log” • % • % gdb a.out log • gdb> run • Program received SIGSEGV. • In get() at para-hash.c:67 • 67 a = bucket->d; Long recording: small log

  14. Our Recorder Desired & Existing Race Recorders Strata ASPLOS ’06 V V V X V V, but global

  15. Small Log Size Coherence Piggyback Transitive Reduction & Regulated TR Order-Value Hybrid Set/LRU Approximation

  16. Problem Formulation Dependence (black) Conflicts (red) Thread I Thread J Thread I Thread J ld A add ld A add st B st B st C st C st C Log st C ld B ld B ld D ld D st A st A sub sub st C st C ld B ld B st D st D Recording Replay • Reproduce exact same conflicts: no more, no less

  17. Dependence Log 1 1 Log J: 23 14 35 46 16 bytes 2 2 3 3 Log I: 23 4 4 5 5 Log Size: 5*16=80 bytes (10 integers) 6 6 Log All Conflicts Thread I Thread J •  Detect conflicts  Write log ld A add st B st C st C ld B ld D st A sub st C ld B st D Replay • Assign IC • (logical Timestamps) • But too many conflicts

  18. TR Reduced Log Log J: 23 35 46 Log I: 23 Log Size: 64 bytes (8 integers) Netzer’s Transitive Reduction Thread I Thread J TR reduced 1 ld A add 1 st B st C 2 2 st C ld B 3 3 ld D st A 4 4 sub st C 5 5 ld B st D 6 6 Replay

  19. From I to J Vectors • Regulate Replay (RTR) From J to I Vectors The Intuition of the New RTR Algorithm After Reduction

  20. New Reduced Log Log J: 23 45 Log I: 23 stricter Reduced Log Size: 48 bytes (6 integers) Stricter Dependences to Aid Vectorization Thread I Thread J 1 ld A add 1 st B st C 2 2 st C ld B 3 3 ld D st A 4 4 sub st C 5 5 ld B st D 6 6 Replay

  21. Vectorized Log Log J: x=3,5, ∆=1 Log I: x=3, ∆=1 Vector Deps. Log Size: 40 bytes (5 integers) Compress Vectorized Dependencies Thread I Thread J 1 ld A add 1 st B st C 2 2 st C ld B 3 3 ld D st A 4 4 sub st C 5 5 ld B st D 6 6 Replay • Reduce log size to KB/core/second

  22. Low Runtime Overhead Coherence Piggyback Transitive Reduction & Regulated TR Set/LRU Approximation Order-Value Hybrid

  23. B.writer = (I, 2) C.writer =(J, 2) if (C.writer != I) log(WAW) foreach C.readers if (reader != I) log(WAR) C.readers.clear( ) C.writer = (I, 3) if (B.writer != J) log(RAW) B.readers.add(J,3) … Detect Conflicts A.readers A.writer Thread I Thread J A.readers.add(I, 1) 1 ld A add 1 st B st C 2 2 st C ld B 3 3 st A 4 Recording • Expensive in software

  24. Get/S Request A.readers A.writer B.readers B.writer Data Response Timestamp Use Cache and Cache Coherence Proc I Proc J ld B Tag State Data Timestamp A S … 1 B M … 2 Tag State Data Timestamp A S … 3 B I … 2 RAW Detected & Logged • Detect conflict in hardware with little runtime cost

  25. Ack Timestamp? Inv Get/S Cache Evictions and Writebacks Proc I Proc J st A Tag State Data Timestamp A S … 1 B M … 2 Tag State Data Timestamp A S … 3 B I … 2 M … 3 C M … 3 WAR Detected & Logged Directory of A: Shared(I,J) Owner() • OK with nonsilent eviction & directory eviction

  26. Implement TR and RTR in Hardware • Ideal TR requires vector timestamps • Too expensive • New idea: Pairwise-TR (use scalar timestamp) • Enable pairwise transitive reduction • Optimal RTR algorithm is likely expensive • Implement a greedy RTR algorithm • One-pass, online algorithm • Keep a sliding window of vectorizable dependencies

  27. Hardware Implementation

  28. Coherence Piggyback Transitive Reduction & Regulated TR Low Cost Hardware Set/LRU Approximation Order-Value Hybrid

  29. C M … 3 Timestamp Approximation Thread I Thread J 1 ld A add 1 One Set of I’s $ Tag State Data Timestamp A S … 1 B M … 2 st B st C 2 2 st C ld B 3 3 Use current IC of thread I I ld D st A J Recording Directory of A: Shared(I) • Correct, but more evictions  more logged conflicts

  30. Hardware Cost Log Size

  31. Thread I Thread J 1 ld A add 1 One Set of I’s $ Tag State Data Timestamp A S … 1 B M … 2 st B st C 2 2 C M … 3 st C ld B 3 3 LRU guarantee B’s TS > A’s TS Use current IC of thread I I ld D st A J Recording Set/LRU Approximation • Set/LRU better preserve reducibility • Small $  more misses  but still small log

  32. Hardware Cost of Timestamps Coupled Timestamp Memory • Coupled timestamp memory: overhead  cache size • Not flexible • 64B line + 64b (24b) timestamp  12.5% (4.7%) overhead • 192 KB for a 4MB L2 • Need to modify cache Tag State Data Timestamp A S … 1 B M … 2

  33. Cache Tag State Data A S … B M … Tag Timestamp A 1 B 2 Timestamp Memory Decoupled Timestamp Memory • Decoupling  Small timestamp memory (Set/LRU) • e.g., 32-set, 64-way  99% transitive reduction • Timestamps Memory  24 KB • No need to modify cache Coupled Timestamp Memory Tag State Data Timestamp A S … 1 B M … 2 • From 192 KB to 24 KB: 8x reduction

  34. Coherence Piggyback Transitive Reduction & Regulated TR Set/LRU Approximation Order-Value Hybrid SC & TSO Applicability

  35. Thread I Thread J A=B=0 st A,1 st A,1 st B,1 ld A A=1 A=0 A=1 A=0 1 st A,1 st B,1 1 st B,1 ld B B=0 B=1 B=0 B=1 ld A ld B ld B st B,1 st A,1 st A,1 ld B ld A 2 2 ld A ld A ld B st B,1 SC TSO Recording with Total Store Order (TSO) • Majority of existing MP are non-SC • TSO is well defined, x86-like

  36. A=0 B=0 TSO Execution I J A=1 B=1 st A,1 st B,1 Thread I Thread J WrBuf WrBuf ld A A=B=0 ld B 1 st A,1 st B,1 1 st A,1 ld B ld A 2 2 Memory System st B,1 A=0 A=0 B=0 B=0

  37. Thread I Thread J 1 st A,1 st B,1 1 ld B ld A 2 2 A=0 Replay B=0 Value Used A=0 Order-Value-Hybrid Recording WAR Omitted Value Logged st A,1 Thread I Thread J I J A=1 B=1 st B,1 A=B=0 ld A 1 st A,1 st B,1 1 WrBuf WrBuf ld B ld B ld A st A,1 2 2 Recording st B,1 Memory System A Changed! A=0 A=0 B=0 B=0 J Starts to Monitor A I Starts to Monitor B I Stops Monitoring B

  38. Hybrid Recording with TR and RTR • Hybrid recording • All loads get correct values • Hardware similar to OoO SC [Gharachorloo et al. ’91] • Hybrid + TR & RTR • TR will not use the omitted WAR in reduction • RTR vectorize dependencies more conservatively

  39. Evaluation Method & Results

  40. Core 4 Core 1 TSM TSM Shared L2 Cache (L1 Dir) IC Core 3 Core 2 L1_I$ L1_D$ L1 Coherence Controller TSM TSM TSM Log TR Reg RTR Reg Put-it-together: Determinizer/CMP

  41. Simulation Method • Commercial server hardware • GEMS: http://www.cs.wisc.edu/gems • Full-system (OS + application) executions • 4-core CMP (Sequential Consistent) • 1-way in-order issue, 2 GHz, • 64KB I/D L1, 4MB L2, 64byte lines, MOSI directory • Commercial server software • Apache – static web serving • SpecJBB – middleware • OLTP – TPC-C like • Zeus – static web serving

  42. KB/core/s byte/core/kilo-instr 200 2.0 150 1.5 100 1.0 50 0.5 0 0.0 Apache JBB OLTP Zeus AVG Apache JBB OLTP Zeus AVG Log Size: 1 byte/kilo-instr • Well within in the capability of current machines • Long recording (days – months) need improvement

  43. Execution Time 100 100 80 80 60 60 40 40 20 20 0 0 Apache JBB OLTP Zeus Apache JBB OLTP Zeus Baseline With race recorder Runtime Overhead Interconnection Msg. B/W • Our recorder can be “always-on”

  44. 100 100 80 80 60 60 40 40 20 20 0 0 Apache JBB OLTP Zeus AVG Apache JBB OLTP Zeus AVG Perfect TSM 24KB Set/LRU TSM Benefits of RTR and Set/LRU (Log Size) Improvement by RTR Effectiveness of Set/LRU Log Size Log Size Pairwise-TR Our RTR

  45. Why RTR and Set/LRU Work Well? • RTR • Processors execute instructions at similar speed • Therefore, we can find “vectorizable” dependencies • Set/LRU • Temporal locality makes the LRU timestamps old • We only need to know if a timestamp is “old-enough”

  46. Sensitivity and Scalability • A design space of the timestamp memory (TSM) • Size: smaller TSM -> larger log • Read/write timestamp: should be used when TSM is large • Partial timestamp: 24-bit enough • Associativity: higher better for RTR • Scalability of the recorder • Studied with modest processors (2p – 16p) • Commercial workloads, not scientific workloads • Log size increase slowly with number of cores

  47. Conclusions, etc.

  48. Conclusions & Future Work • Race recording  Key to combat nondeterminism • Contributions  Effective & inexpensive Recorder • Transitive Reduction & RTR algorithm small log size • Coherencepiggyback Negligible slowdown • Timestamp approximation Low hardware cost • Order-value hybrid  support SC & TSO • Future work • Operate with Hardware Transactional Memory • Seek to Eliminate Timestamp on Acknowledgements

  49. Pull Shared Get/X Toward Recording w/ Snooping Protocols • Key problem is combined/implicit response • Not a problem for AMD Hammer Proc I Proc J st A Tag State Data Timestamp A S … 1 B M … 4 Tag State Data Timestamp A S … 3 B I … 2 + Current IC WAR Detected & Logged

  50. Ack Timestamp Eviction Get/S Timestamp Memory Timestamp at L2-Directory or Memory? Proc I Proc J st A Tag State Data Timestamp A S … 1 B M … 4 Tag State Data Timestamp A S … 3 B I … 2 M … 4 C M … 3 Directory of A: Shared(J) Owner() StickyS(I,J) • Directory eviction: more false conflict, like snooping

More Related