Multi-view data types Scalable concurrency in the multi-core era - PowerPoint PPT Presentation

Multi-view data types Scalable concurrency in the multi-core era Deepthi Akkoorath 1 , Jos´ ao 2 , Annette Bieniusa 1 , Carlos Baquero 3 e Brand˜ 1 Technical University of Kaiserslautern Germany 2 Universidade do Minho Braga, Portugal 3 HASLab, Universidade do Minho & INESC TEC Braga, Portugal

Concurrent programs in multi-core Speedup Synchronization Threads 2/13

Overview Distributed systems • Eventual consistency + CRDTs � Synchronisation free • Fast, Scalable, Available Goal • Weak consistency � Less synchronisation • Speed up! 3/13

Overview Distributed systems • Eventual consistency + CRDTs � Synchronisation free • Fast, Scalable, Available Goal • Weak consistency � Less synchronisation • Speed up! Global-Local view model Multi-view data types 3/13

Global-local view model Thread A Thread B Local view 10 10 10 Global view 4/13

Global-local view model Thread A Thread B c.inc() Local view 11 10 10 Global view 4/13

Global-local view model Thread A Thread B c.inc() c.inc() Local view 11 11 Global view 10 4/13

Global-local view model Thread A Thread B c.inc() c.inc() c.inc() Local view 11 12 Global view 10 4/13

Global-local view model Thread A Thread B c.inc() c.inc() c.inc() c.merge() 11 12 Local view Global view 12 4/13

Global-local view model Thread A Thread B c.inc() c.inc() c.merge() c.inc() c.merge() Local view 13 12 13 Global view 4/13

Operations Local View: (snapshot, local) Global view that was Accumulates local last seen by this thread updates Global View: g 5/13

Operations weakUpdate weakRead Local View: (snapshot, local) Global view that was Accumulates local last seen by this thread updates Global View: g 5/13

Operations weakUpdate weakRead Local View: (snapshot, local) Global view that was Accumulates local last seen by this thread updates Global View: g strongUpdate strongRead 5/13

Operations weakUpdate weakRead Local View: (snapshot, local) Global view that was Accumulates local last seen by this thread updates merge Global View: g strongUpdate strongRead 5/13

Multi-view data types Mergeable types • Implements weak operations and merge Hybrid types • Implements weak, strong and merge operations • Hybrid counter synchronous increment when close to a target • Hybrid queue weak enqueue and synchronous dequeue 6/13

CRDTs? 7/13

CRDTs? • G-Set • merge = union of sets • Counter • Map: id → int • merge = max of each elem 7/13

CRDTs? • G-Set • merge = union of sets • Counter • Map: id → int • merge = max of each elem CRDT merge is expensive 7/13

CRDTs? • G-Set • merge = union of sets • Counter • Map: id → int • merge = max of each elem CRDT merge is expensive Multi-view data types • Multiple versions (view) • Isolated access to each view • Fast merge 7/13

Counter weakInc ( ) { l ++; } • Global view: int g weakValue ( ) { • Local view : return s+ l ; ( int s, int l ) } merge ( ) { Thread-local copies atomic { g += l ; Exclusive access ⇒ no s = g ; l = 0; } synchronization } Synchronous merge strongInc ( ) { atomic { g++; } } 8/13

Multi-view list T1 head T2 After T 1 commits: T1 head T2 After T 2 commits: T1 head T2 9/13

Evaluation: Hybrid Counter Goal: increment until a target Periodic merge ⇒ Divergence from target Switches to strong update after a threshold � ������ � � � �� � � ��� �� � ��� � ���� � ��� � ��� ���� � ��� ����� � ��� ���� � ������ � ������� � ���� � � ���� � � � � � � � � � � �� � �� � �� � �� � �� � �� � �� ������� 10/13

Evaluation: Breadth first traversal Using hybrid queue : weak enqueue and strong dequeue 4.5 4 3.5 3 Speed up 2.5 2 1.5 Lin-Lockbased Lin-LockFree 1 M-Lockbased M-LockFree 0.5 1 2 4 8 12 16 20 24 No.of threads 11/13

Related work Mergeable types Weak consistency Doppel [Narula et al., 2014] Quasi linearizability [Afek et al., 2010] in-memory transactions Weak/medium future linearizability [Kogan and Herlihy, 2014] Concurrent revisions K-linearizability [Aiyer et al., 2005] [Burckhardt et al., 2010] Quiescent consistency fork join model [Aspnes et al., 1994] “mergeable” types 12/13

Summary Global-local view model • fast local state, distant global state Impact on underlying data structure design • Multiple versions, Merge Combination of weak and strong updates • A spectrum of consistency Thank you! akkoorath@cs.uni-kl.de 13/13

References I Afek, Y., Korland, G., and Yanovsky, E. (2010). Quasi-linearizability: Relaxed consistency for improved concurrency. In Proceedings of the 14th International Conference on Principles of Distributed Systems , OPODIS’10, pages 395–410, Berlin, Heidelberg. Springer-Verlag. Aiyer, A., Alvisi, L., and Bazzi, R. A. (2005). On the availability of non-strict quorum systems. In Proceedings of the 19th International Conference on Distributed Computing , DISC’05, pages 48–62, Berlin, Heidelberg. Springer-Verlag. Aspnes, J., Herlihy, M., and Shavit, N. (1994). Counting networks. J. ACM , 41(5):1020–1048. Burckhardt, S., Baldassin, A., and Leijen, D. (2010). Concurrent programming with revisions and isolation types. In Proceedings of the ACM International Conference on Object Oriented Programming Systems Languages and Applications , OOPSLA ’10, pages 691–707, New York, NY, USA. ACM. 1/2

References II Kogan, A. and Herlihy, M. (2014). The future(s) of shared data structures. In Proceedings of the 2014 ACM Symposium on Principles of Distributed Computing , PODC ’14, pages 30–39, New York, NY, USA. ACM. Narula, N., Cutler, C., Kohler, E., and Morris, R. (2014). Phase reconciliation for contended in-memory transactions. In Proceedings of the 11th USENIX Conference on Operating Systems Design and Implementation , OSDI’14, pages 511–524, Berkeley, CA, USA. USENIX Association. 2/2