how to implement any concurrent data structure

Transcription

1 how to implement any concurrent data structure marcos k. aguilera vmware jointly with irina calciu siddhartha sen mahesh balakrishnan

2 Where to find more information about this work How to Implement Any Concurrent Data Structure. By Irina Calciu, Siddhartha Sen, Mahesh Balakrishnan, Marcos K. Aguilera. Communications of the ACM, 2018 Black-box Concurrent Data Structures for NUMA Architectures. Irina Calciu, Siddhartha Sen, Mahesh Balakrishnan, Marcos K. Aguilera. ASPLOS, 2017

3 concurrent data structures are everywhere kernel application libraries applications

4 but efficient ones are hard to design locks transactional memory lock-free and wait-free

5 effort in The Future(s) of Shared Data Structures Alex Kogan and Maurice Herlihy PODC 2014 Concurrent Updates with RCU: Search Tree as an Example Maya Arbel and Hagit Attiya PODC 2014 Dynamic-Sized Nonblocking Hash Tables Yujie Liu, Kunlong Zhang and Michael Spear PODC 2014 Efficient Lock-free Binary Search Trees Bapi Chatterjee, Nhan Nguyen and Philippas Tsigas PODC 2014 The Amortized Complexity of Non-blocking Binary Search Trees Faith Ellen, Panagiota Fatourou, Joanna Helga and Eric Ruppert PODC 2014 The Adaptive Priority Queue with Elimination and Combining Irina Calciu, Hammurabi Mendes and Maurice Herlihy DISC 2014 Solo-fast Universal Constructions for Deterministic Abortable Objects Claire Capdevielle, Colette Johnen and Alessia Milani DISC 2014 On Deterministic Abortable Objects Vassos Hadzilacos and Sam Toueg PODC 2013 Leaplist: Lessons Learned in Designing TM-Supported Range Queries Hillel Avni, Nir Shavit, and Adi Suissa PODC 2013 The SkipTrie: Low-Depth Concurrent Search without Rebalancing Rotem Oshman and Nir Shavit PODC 2013 Pragmatic Primitives for Non-blocking Data Structures Trevor Brown, Faith Ellen, and Eric Ruppert PODC 2013 Lock-Free Data Structure Iterators Erez Petrank and Shahar Timnat DISC 2013 Practical Non-blocking Unordered Lists Kunlong Zhang, Yujiao Zhao, Yajun Yang, Yujie Liu and Michael Spear DISC 2013 Atomic snapshots in expected $O(\log^3 n)$ steps using randomized helping James Aspnes and Keren Censor-Hillel DISC 2013 An Optimal Implementation of Fetch-and-Increment Faith Ellen and Philipp Woelfel DISC 2013 On the Time and Space Complexity of Randomized Test-And-Set George Giakkoupis and Philipp Woelfel PODC 2012 Universal Constructions that Ensure Disjoint-Access Parallelism and Wait- Freedom Faith Ellen, Panagiota Fatourou, Eleftherios Kosmas, Alessia Milani, and Corentin Travers PODC 2012 Faster than Optimal Snapshots (for a While) James Aspnes, Hagit Attiya, Keren Censor-Hillel, and Faith Ellen PODC 2012 Strongly Linearizable Implementations: Possibilities and Impossibilities Maryam Helmi, Lisa Higham, and Philipp Woelfel PODC 2012 CBTree: A Practical Concurrent Self-Adjusting Search Tree Yehuda Afek, Haim Kaplan, Boris Korenfeld, Adam Morrison, Robert E. Tarjan DISC 2012 Efficient Fetch-and-Increment Faith Ellen, Vijaya Ramachandran, Philipp Woelfel DISC 2012

6 problems with concurrent data structure design herculean effort for each data structure rigid designs an even greater problem

7 problems with concurrent data structure design herculean effort for each data structure rigid designs an even greater problem new hardware architectures

8 our options? 1. underutilize the system 2. develop new data structures for each new architecture 3. we think there is a better way

9 architecture-aware black-box data structures transformation 1 architecture 1 sequential data structures transformation 2 architecture 2 transformation 3 architecture 3

10 architecture-aware black-box data structures FOCUS OF REST OF TALK transformation 1 NUMA architecture architecture 1 sequential data structures transformation 2 architecture 2 transformation 3 architecture 3

11 the NR algorithm

12 NUMA architecture Non-Uniform Memory Access node node core core core core core core core core cache cache cache cache cache cache cache cache cache cache memory memory local access more efficient

13 evaluation Intel Xeon E7-4850v3 56 cores, 4 nodes 2.2 GHz 512 GB RAM L3 35 MB L2 256 KB L1 64 KB

14 skip list priority queue 10% updates X (NR) Node Replication (FC+) FC + RWL (RWL) Readers-Writer Lock (LF) Lock-free X (FC) Flat Combining (SL) Spinlock 60 ops/us # threads

15 data structure in REDIS: 10% updates X (NR) Node Replication (FC+) FC + RWL (RWL) Readers-Writer Lock X (FC) Flat Combining (SL) Spinlock 6 ops/us # threads

16 the transformation given single-threaded execute(op,parameters) isreadonly(op) we produce multi-threaded execute(op,parameters) works well in NUMA servers

17 key ideas 1. replicate data structure across (NUMA) nodes state machine approach with a shared log 2. provide efficient NUMA-aware log large effort to optimize log

18 the transformation NUMA Node Local Replica NUMA Node Local Replica Thread Thread Thread Thread

19 the transformation NUMA Node Local Replica Shared Log NUMA Node Local Replica Local Tail Local Tail Thread Thread LogTail Thread Thread

20 how to implement log? key observation coordination within node cheaper than across nodes within node: we use flat combining across nodes: we use lock-free appending to log

21 correctness linearizability [Herlihy Wing 1990]: each operation appears to take effect instantaneously at a point between its invocation and response

22 whence performance comes trade memory + computation for less communication compact representation of operations limited cross-node synchronization and contention enable parallelism combiners across nodes readers within a node readers and the combiner on the same node leverage batching 22

23 conclusion fundamental changes in hardware exposed to software developers take-away: instead of individual data structures, let s develop general architecture-aware techniques