Transac'onal Libraries Alexander Spiegelman *, Guy Golan-Gueta, and Idit Keidar * Technion Yahoo Research

Transcription

1 Transac'onal Libraries Alexander Spiegelman *, Guy Golan-Gueta, and Idit Keidar * * Technion Yahoo Research 1

2 Mul'-Threading is Everywhere 2

3 Agenda Concurrent Data Structure Libraries (CDSLs) vs Memory Introducing: Data Structure Libraries (TDSL) Example TDSL algorithm Skiplist of objects (skiplists and queues) Fast abort-free singletons Library 3

4 Data Structures (DS) building blocks in modern SW Map, skiplist, queue, etc. 4

5 But Are They Thread Safe? Correct under concurrency? OR? 5

6 Thread-Safe Concurrent DS Libraries Widely used in real life sosware Numerous research papers Concurrent Skipklist [Herlihy, Lev, Luchangco, Shavit: A simple op@mis@c skiplist algorithm] [Fraser: Prac@cal lock freedom] Concurrent queue [Michael, Sco[: Simple, fast, and prac@cal nonblocking and blocking concurrent queue algorithms] [Gramoli, Guerraoui: Reusable concurrent data types] Concurrent binary tree [Bronson, Casper, Chafi, Olukotun: A prac@cal concurrent binary search tree] [Drachsler, Vechev, Yahav: Prac@cal concurrent binary search trees via logical ordering] 6

7 Concurrent Data Structure Libraries (CDSLs) Each executes atomically Custom-tailored preserves Thread Thread Thread map queue 7

8 Are They Really Thread Safe?? balance ßmap.get (key=yoni) newbalanceß balance+deposit map.set(key=yoni,newbalance) repq.enq( Yoni s balance=new ) balance ßmap.get(key=Yoni) newbalanceß balance+deposit map.set(key=yoni,newbalance) repq.enq( Yoni s balance=new ) Oops! Atomic opera@ons are not enough 8

9 SoGware Transac'onal Memory (STM) TL2, TinySTM, SwissTM, (TXs) = atomic sec@ons including mul@ple DS opera@ons Begin_TX balanceßmap.get(key=yoni) newbalance ß balance + deposit map.set(key=yoni, newbalance) repq.enq( Yoni s balance=new ) End_TX begin end abort App map STM queue 9

10 STM Overhead Explained Common STM Each object has a version TX reads a consistent snapshot Validates versions have not changed by Otherwise aborts TX updates occur during commit Sources of overhead: Global version clock (GVC) high conten@on Read- and write-sets tracking and valida@on overhead Conflicts aborts 10

11 CDSL vs STM CDSL STM Performance ü û Exploit DS & structures ü û Used in ü û Generality û ü Composability û ü 11

12 Agenda Concurrent Data Structure Libraries (CDSLs) vs Memory Introducing: Data Structure Libraries (TDSL) Example TDSL algorithm Skiplist of objects (skiplists and queues) Fast abort-free singletons Library 12

13 TDSL: Bringing Transac'ons into CDSL Legacy singleton fast, abort-free App App App begin end abort STM programmability: TXs span any number of Custom-tailored, performance of CDSL map queue Library Missing: Generality of STM 13

14 Why TDSL? Power of using DS and structure Use to reduce overhead aborts E.g., read-set index Different CC mechanisms for different DSs maps, sets queues Fast abort-free singletons As fast as in CDSL No on global version clock Support legacy code 14

15 TDSL Benefit 1: Programmability Support for legacy code Fast, abort-free singletons Power of Begin_TX valßmap.get(key=yoni) newß val+deposit map.set(key=yoni,new) repq.enq( Yoni s balance=new ) End_TX App App App map queue Transac@onal Library 15

16 TDSL Benefit 2: Seman'c Op'miza'on Use known But, take advantage of & structure of each DS to reduce aborts & overhead Reduce read-set Do some of the work 16

17 Example of Abort Reduc'on Two concurrent put Put(23) traverses the list, reads nodes that put(4) updates Conflict, STM would abort at least one But no conflict

18 Example of Abort Reduc'on Two concurrent put Put(23) traverses the list, reads nodes that put(4) updates Conflict, STM would abort at least one But no conflict

19 Example of Abort Reduc'on Two concurrent put Put(23) traverses the list, reads nodes that put(4) updates Conflict, STM would abort at least one But no conflict

20 TDSL Benefit 3: Mix & Match Maps allow lots of concurrency Amenable to Queues do not be[er STM picks one Our TDSL can mix & match App App App map queue Library 20

21 Agenda Concurrent Data Structure Libraries (CDSLs) vs Memory Introducing: Data Structure Libraries (TDSL) Example TDSL algorithm Skiplist of objects (skiplists and queues) Fast abort-free singletons Library 21

22 Skiplist Roadmap 1. Add STM-like support to simple linked list Based on TL2 STM algorithm 2. remove redundant and tracking Reduce aborts Use and structure 3. shorten Reduce aborts index Lazy GC 22

23 Step 1: Standard STM Take a simple linked list Add TL2 mechanism to support TXs: Add a version to each node Get versions from global version clock (GVC) Maintain read-set with read versions Defer updates, track in write-set To commit: lock write-set, validate read-set, increment GVC, update 23

24 Step 1: Standard STM insert(10): Local memory Shared memory 10 Read read-set Validate write-set GVC

25 Step 2: Reduce Read-Set Exploit structure and insert(10): 10 read-set 9 write-set 9 10 Local memory Shared memory Read Validate GVC

26 Step 3: Non-Transac'onal Index Imagine we could guess the right node GVC So, we could be faster and save aborts during the traversal 26

27 Step 3: Non-Transac'onal Index getsmaller Returns some node with a smaller key For performance, not much smaller Implemented as (concurrent) skiplist For example App Index 27

28 Step 3: Non-Transac'onal Index insert(10): 10 App Start traverse from 5 Index GVC

29 Step 3: Non-Transac'onal Index Updated outside the TX (if completes successfully) Reduces aborts, overhead But: May return nodes with smaller keys than predecessor longer traversals May return removed nodes Subtle, more details in the paper 29

30 Agenda Concurrent Data Structure Libraries (CDSLs) vs Memory Introducing: Data Structure Libraries (TDSL) Example TDSL algorithm Skiplist of objects (skiplists and queues) Fast abort-free singletons Library 30

31 Composi'on One GVC shared among all objects Commit: Lock all write sets Validate all read sets Increase GVC Update objects and release locks 31

32 Agenda Concurrent Data Structure Libraries (CDSLs) vs Memory Introducing: Data Structure Libraries (TDSL) Example TDSL algorithm Skiplist of objects (skiplists and queues) Fast abort-free singletons Library 32

33 Fast Abort-Free Singletons Use the index Do not increment GVC No As fast as in CDSL Make aware of singletons using designated fields Details in the paper 33

34 Agenda Concurrent Data Structure Libraries (CDSLs) vs Memory Introducing: Data Structure Libraries (TDSL) Example TDSL algorithm Skiplist of objects (skiplists and queues) Fast abort-free singletons Library 34

35 Composing TDSLs Extend API to support: TX-begin Divide TX-commit into three phases: TX-lock, TX-verify, and TX-finalize TX-abort Perform TX-begin and each of the commit phase in all libraries Wait for in all libraries before moving to the next phase If catch abort in some library, perform TX-abort in all Some have to be Based on theory of PLDI 15 paper by Ziv et al. [Ziv, Aiken, Golan-Gueta, Ramalingam, Sagiv. Composing concurrency control] 35

36 Compose With Exis'ng Libraries Modify library to support the extanded API E.g., TL2 STM naturally supports it E.g., use standard 2 phase locking Together we get general transac@ons Ones supported fully by a custom-tailored TDSL are fast Others less so 36

37 Agenda Concurrent Data Structure Libraries (CDSLs) vs Memory Introducing: Data Structure Libraries (TDSL) Example TDSL algorithm Skiplist of objects (skiplists and queues) Fast abort-free singletons Library 37

38 MILLIONS OPS/SEC READ-ONLY WORKLOAD # OF THREADS Singletons UPDATE-ONLY WORKLOAD As good as baseline # OF THREADS our skiplist rota@ng nohotspot fraser op@mis@c baseline Do not support transac@ons 38

39 Singletons UPDATE-ONLY WORKLOAD MILLIONS OPS/SEC # OF THREADS our skiplist rotai@ng nohotspot fraser lock base line 39

40 We have less overhead Transac'ons We have less aborts MILLIONS TXS/SEC READ-ONLY WORKLOAD # OF THREADS UPDATE-ONLY WORKLOAD # OF THREADS our skiplist SeqTL2 FriendlyTL2 40

41 Transac'ons MILLIONS TXS/SEC UPDATE-ONLY WORKLOAD # OF THREADS our skiplist seq TL2 friendly TL2 41

42 Transac'ons (Aborts) % OF ABORTS UPDATE-ONLY WORKLOAD # OF THREADS ZOOMED IN # OF THREADS our skiplist SeqTL2 FriendlyTL2 42

43 Intruder skiplists and queues in a real applica@on SPEED UP # OF THREADS % OF ABORTS Our library SeqTL2 FriendlyTL # OF THREADS cannot support intruder 43

44 onclusion We introduced a new concept for concurrent programing Composable DSs suppor@ng TX as well as fast singletons Implemented library with two DSs Map (based on skiplist) and queue We hope that the community will adopt this concept Build and use more such libraries 44

45 Related Work and [M. Herlihy and E. Koskinen: methodology for highlyconcurrent objects] [Ahmed Hassan, Roberto Palmieri, Peluso, and Binoy Ravindran: Open [Y. Ni, V. S. Menon, A.-R. Adl-Tabatabai, A. L. Hosking, R. L. Hudson, J. E. B. Moss, B. Saha, and T. Shpeisman. Open in sosware memory] Reagents [A. Turon. Reagents: expressing and composing fine-grained concurrency] 45

46 Conclusion CDSL STM TDSL Performance ü û ü Exploit DS ü û ü Used in ü û Generality û ü ü Composability û ü ü We hope that the community will adopt this concept Build and use more such libraries 46