the universal transactional memory construction
play

THE UNIVERSAL TRANSACTIONAL MEMORY CONSTRUCTION Jons-Tobias - PowerPoint PPT Presentation

THE UNIVERSAL TRANSACTIONAL MEMORY CONSTRUCTION Jons-Tobias Wamhoff and Christof Fetzer Dresden University of Technology, Germany 1 MOTIVATION Universal construction Shows how to convert sequential algorithm into concurrent wait-free


  1. THE UNIVERSAL TRANSACTIONAL MEMORY CONSTRUCTION Jons-Tobias Wamhoff and Christof Fetzer Dresden University of Technology, Germany 1

  2. MOTIVATION • Universal construction • Shows how to convert sequential algorithm into concurrent wait-free algorithm • Can one base such a construction on TM? • Wait-free progress for all correct operations • Tolerate crashes and non-terminating operations The Universal Transactional Memory Construction 2 TRANSACT‘11, San Jose, CA

  3. ASYNCHRONOUS MULTICORE SYSTEM MODEL (AMSM) • Asynchronous model with features of Shared Memory current multi-core systems Transactional Memory • Performance counters for executed cycles per thread Code Code • Size of memory is bound • Operations (transactions) can be Perf Ctr 1 Perf Ctr n executed by any thread Thread 1 Thread n • Compare-and-Swap (CAS), Operating System Fetch-and-Increment (FAI) The Universal Transactional Memory Construction 3 TRANSACT‘11, San Jose, CA

  4. AMSM CRASH FAILURES Shared Memory • Threads can crash (stop taking steps) Transactional Memory caused by: • Operating system, hardware, signal: Code Code not detectable in AMSM Perf Ctr 1 Perf Ctr n • Program code (bug): detected by Thread 1 Thread n runtime and converted in exception Operating System The Universal Transactional Memory Construction 4 TRANSACT‘11, San Jose, CA

  5. AMSM PROGRAMS • Log defines total order on a sequence of Log 0 Log 1 invocations (operations) S1 S2 Invocation 0 • FIFO sequential execution txBegin txBegin txRead txRead • Multiple logs executed in parallel txWrite txWrite txCommit txCommit • Invocations (transactions) applied to shared memory S3 S4 Invocation 1 • Sequential object to transform into wait-free txBegin txBegin linerizable object txRead txRead txWrite txWrite • Transactions can be non-terminating txCommit txCommit • Terminate correct transactions within finite number of steps The Universal Transactional Memory Construction 5 TRANSACT‘11, San Jose, CA

  6. INVOCATION PROCESSING Log 0 Log 1 Shared Memory S1 S2 Invocation 0 Transactional Memory txBegin txBegin txRead txRead How can one txWrite txWrite txCommit txCommit process invocations Code Code from k logs using n S3 S4 threads in a finite Invocation 1 txBegin txBegin number of steps? Perf Ctr 1 Perf Ctr n txRead txRead txWrite txWrite Thread 1 Thread n txCommit txCommit Operating System The Universal Transactional Memory Construction 6 TRANSACT‘11, San Jose, CA

  7. UNIVERSAL TRANSACTIONAL MEMORY CONSTRUCTION • Universal construction transforms program from one valid state to another valid state • Use TM to isolate modifications until commit • Schedule transactions on threads for wait-free progress The Universal Transactional Memory Construction 7 TRANSACT‘11, San Jose, CA

  8. STATES current Memory State History ... S0 S1 S2 S3 NULL chunks 0 0 0 1 0 1 0 3 1 0 1 0 1 2 1 2 2 0 2 0 2 2 2 2 3 0 3 1 3 1 3 3 4 0 4 0 4 0 4 3 key ct ... The Universal Transactional Memory Construction 8 TRANSACT‘11, San Jose, CA

  9. TRANSACTION SCHEDULING ONTO THREADS Log 1 ... Log k S1 Sk • A thread‘s commit time is FAI( CT ) Invocation 0 txBegin txBegin txRead txRead • S CT = f tx ( S CT -1 ) txWrite txWrite txCommit txCommit • CT -1 = i * k + l Sk+1 S2k Invocation 1 txBegin txBegin • Log l = ( CT -1) mod k txRead txRead txWrite txWrite txCommit txCommit • Invocation i = ( CT -1) div k The Universal Transactional Memory Construction 9 TRANSACT‘11, San Jose, CA

  10. TRANSACTION SCHEDULING ONTO THREADS Log 0 Log 1 S1 S2 • ct 4 = FAI( CT 4 ) Invocation 0 txBegin txBegin txRead txRead • S 4 = f tx ( S 3 ) txWrite txWrite txCommit txCommit • ct 4-1 = i * k + l S3 S4 Invocation 1 txBegin txBegin • Log l 1 = 3 mod 2 txRead txRead txWrite txWrite txCommit txCommit • Invocation i 1 = 3 div 2 The Universal Transactional Memory Construction 10 TRANSACT‘11, San Jose, CA

  11. txBegin(ct) current • Clone current head of state history S4 S2 NULL B2 • Chunks are contained only by 0 1 0 1 reference 1 2 1 2 2 2 2 2 • Set new commit time 3 1 3 1 4 0 4 0 • Keep base version key ct The Universal Transactional Memory Construction 11 TRANSACT‘11, San Jose, CA

  12. txWrite(addr, val) current S4 S2 NULL B2 • Identify chunk using hash function 0 1 0 1 • Clone chunk if still reference 1 2 1 4 2 2 2 2 • Update clone with value 3 1 3 1 4 0 4 0 key ct The Universal Transactional Memory Construction 12 TRANSACT‘11, San Jose, CA

  13. txRead(addr) current S4 S2 • Identify chunk using hash function NULL B2 • Keep commit time of chunk in read- 0 1 0 1 1 2 1 4 set if written by predecessor 2 2 2 2 3 1 3 1 • Return value of address 4 0 4 0 key ct The Universal Transactional Memory Construction 13 TRANSACT‘11, San Jose, CA

  14. txCommit • Wait until S CT -1 is available current S4 S2 S3 • Validate against S CT -1 that read- B2 set versions unchanged 0 1 0 3 0 1 1 2 1 2 1 4 • Failure : abort and retry 2 2 2 2 2 2 3 1 3 3 3 1 • Success : update chunk 4 0 4 3 4 0 references and append state to history using CAS key ct The Universal Transactional Memory Construction 14 TRANSACT‘11, San Jose, CA

  15. txCommitProceeding(ct) current S2 S3 S4 • Wait until S CT -1 is available 0 1 0 3 0 3 • Clone S CT -1 and append as S CT 1 2 1 2 1 2 2 2 2 2 2 2 using CAS 3 1 3 3 3 3 4 0 4 3 4 3 key ct The Universal Transactional Memory Construction 15 TRANSACT‘11, San Jose, CA

  16. UNIVERSAL CONSTRUCTION FOR THE GOOD CASE Log 0 Log 1 Log 2 Global State s1=ftx(S0) s2=ftx(S1) s3=ftx(S2) S3 Invocation 0 s4=ftx(S3) s5=ftx(S4) s6=ftx(S5) Invocation 1 NULL NULL NULL Invocation 2 next CT CT8 ct4 ct5 ct6 ct7 Thread 1 Thread 2 Thread 3 Thread 4 Operating System The Universal Transactional Memory Construction 16 TRANSACT‘11, San Jose, CA

  17. DEALING WITH THREAD CRASHES Log 0 Log 1 S3 S4 • Thread crashes not detectable in Invocation 1 AMSM: need helping txBegin txBegin txRead txRead txWrite txWrite txCommit txCommit • Waiting in txCommit for S CT -1 used for helping ct3 ct4 • Process S CT -1 = f tx ( S CT -2 ) Thread 1 Thread 2 current • Progress as long as one thread survives S2 S4 NULL The Universal Transactional Memory Construction 17 TRANSACT‘11, San Jose, CA

  18. DEALING WITH TRANSACTION CRASHES Log 0 Log 1 S3 S4 • Transaction crashes detectable in Invocation 1 AMSM: throw exception txBegin txBegin txRead txRead txWrite txWrite txCommit txCommit • Considered as persistent failures Invocation 2 NULL S5=S4 • Commit proceeding state Empty • No further invocations can be added current to the log S2 S3=S2 NULL The Universal Transactional Memory Construction 18 TRANSACT‘11, San Jose, CA

  19. TOLERATING NON- TERMINATING TRANSACTIONS Log 0 Log 1 S3 S4 t=12 t=23 • Non-terminating transaction: f tx ( S CT -1 ) never returns Invocation 1 txBegin txBegin txRead txRead • Use performance counter to assign quota of steps txWrite txWrite for invocation txCommit txCommit • When exceeded commit proceeding and retry NULL S5 t=13 with larger quota Invocation 2 txBegin • Quota is unknown and usually large txRead txWrite txCommit • Bounded number of states (bound memory) means bound number of steps for transaction to current complete S2 S3=S2 NULL The Universal Transactional Memory Construction 19 TRANSACT‘11, San Jose, CA

  20. CONCLUSION • Wait-free progress under AMSM by: • Decoupling of work from threads for helping • Isolating changes in transactions to mask failures The Universal Transactional Memory Construction 20 TRANSACT‘11, San Jose, CA

Recommend


More recommend