Fast Bounded-Concurrency Hash Tables. Samy Al Bahra

Transcription

1 Fast Bounded-Concurrency Hash Tables Samy Al Bahra /

2 Introduction A general mechanism for achieving non-blocking progress guarantees cheaply. Support for any open-addressed collision resolution mechanism. No atomic operations, memory barriers or memory allocation on the fast path (for TSO). Writers never block. Readers only block for a small subset of executions. Dead simple.

3 Motivation Single-writer multi-reader (SWMR) workload. Constant stream of read operations but bursts of millions of writes every few minutes. Firm real-time requirements on read-side. Starvation-freedom on write-side. Systems were memory-constrained so memory efficiency is important.

4 Motivation State-of-the-art entailed significant trade-off in complexity and performance.

5 Motivation State-of-the-art entailed significant trade-off in complexity and performance. Chaining Memory management. Reliance on expensive operations. Limitations on collision resolution mechanism.

6 Motivation State-of-the-art entailed significant trade-off in complexity and performance. Chaining Memory management. Reliance on expensive operations. Limitations on collision resolution mechanism. Open Addressing Complex object re-use constraints in absence of key duplication. Reliance on expensive operations. Limitations on collision resolution mechanism.

7 Motivation State-of-the-art entailed significant trade-off in complexity and performance. Chaining Memory management. Reliance on expensive operations. Limitations on collision resolution mechanism. Open Addressing Complex object re-use constraints in absence of key duplication. Reliance on expensive operations. Limitations on collision resolution mechanism. What does specialization get us?

8 Constraints Termination-safety is not a requirement. SWMR allows us to rely on less complex instructions. Primarily executing on x86 processors. Load to load ordering. Store to store ordering.

9 Implementation ck_pr_store_x(a, b) atomically { *a = b } ck_pr_load_x(a) atomically { *a } ck_pr_fence_store() smp_wb() ck_pr_fence_load() smp_rb()

10 Implementation

11 delete Implementation

12 Implementation delete insert

13 get Implementation

14 Implementation get search

15 Correctness

16 Correctness

17 Correctness

18 Correctness

19 Correctness

20 Correctness

21 Correctness

22 Observed (D, E, v, D) Correctness

23 Observed (D, E, v, D) Correctness

24 Observed (D, E, v, D) Correctness

25 Observed (D, E, v, D) Correctness

26 Observed (D, E, v, D) Correctness

27 Correctness Cannot observe (D, E, v, D)

28 Correctness Cannot observe (D, E, v, D)

29 Correctness Cannot observe (D, E, v, D)

30 Correctness Cannot observe (D, E, v, D)

31 Correctness Cannot observe (D, E, v, D)

32 Correctness If (k, v ) is observed, then so is D and D.

33 Correctness If (k, v ) is observed, then so is D and D.

34 Correctness If (k, v ) is observed, then so is D and D.

35 Correctness If (k, v ) is observed, then so is D and D.

36 Correctness If (k, v ) is observed, then so is D and D.

37 What about more sophisticated collision resolution techniques?

38 Probe Sequence Mutation

39 Probe Sequence Mutation Insert new key-value pair.

40 Probe Sequence Mutation Insert new key-value pair.

41 Probe Sequence Mutation Delete old key-value pair.

42 Probe Sequence Mutation Delete old key-value pair.

43 Probe Sequence Mutation Delete old key-value pair.

44 Probe Sequence Mutation Delete old key-value pair.

45 Probe Sequence Mutation Delete old key-value pair.

46 Probe Sequence Mutation Delete old key-value pair.

47 Further Specialization Sacrifice write-side progress guarantees for read-side wait-freedom.

48 Recent Updates

49 The End Write-up at: Implementations at: Thanks to Maged Michael, Mathieu Desnoyers, Olivier Houchard, Paul Khuong, Paul McKenney, Theo Schlossnagle, Wez Furlong and the Backtrace team for feedback on article.