object_threadsafe
object_threadsafe copied to clipboard
AlexeyAB
Make get_or_set_index immune to memory address reuse
If a mutex instance gets used, deleted and then recreated in the same thread - especially on the stack, the new instance might get the same memory address as the old instance. Because of this, a lookup in thread_local_index_hashmap can't be based on the instance address (this), but rather on the (heap allocated, shared) shared_locks_array instance address, because that one will not be reused even after the mutex instance is deleted since all unregister_t structs still hold a std::shared_ptr to it until they get erased from thread_local_index_hashmap.
I only ran the benchmark from object_threadsafe/benchmark, hope that's fine, I would expect the fix to have no impact on performance:
Before the change:
CPU Cores: 8
Benchmark thread-safe associative containers with size = 100000
Threads = 8, iterations per thread = 2000000
Time & MOps - steady_clock is_steady = 1, num/den = 1 / 1000000000
Latency - high_resolution_clock is_steady = 0, num/den = 1 / 1000000000
burn_cpu() for each multithread operations took: 3.26855 nano-sec
Filling of containers... filled containers.
0 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 9.02 1.77
std::map & std::mutex: 20.8 0.77
safe_ptr<map,mutex>: 20.4 0.785
safe_ptr<map,shared>: 2.29 6.99
safe_ptr<map,contfree>: 1.3 12.3
safe<map,contf>rowlock: 1.35 11.8
safe part<mutex>: 3.83 4.18
safe part<contfree>: 1.51 10.6
15 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 5.35 2.99
std::map & std::mutex: 14.4 1.11
safe_ptr<map,mutex>: 17 0.944
safe_ptr<map,shared>: 29.5 0.543
safe_ptr<map,contfree>: 3.42 4.68
safe<map,contf>rowlock: 2.68 5.96
safe part<mutex>: 3.22 4.97
safe part<contfree>: 1.59 10
30 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 4.48 3.58
std::map & std::mutex: 15.4 1.04
safe_ptr<map,mutex>: 14.2 1.13
safe_ptr<map,shared>: 35.6 0.449
safe_ptr<map,contfree>: 4.83 3.32
safe<map,contf>rowlock: 3.78 4.24
safe part<mutex>: 3.7 4.32
safe part<contfree>: 1.74 9.21
45 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 4.57 3.5
std::map & std::mutex: 14.7 1.09
safe_ptr<map,mutex>: 13.6 1.18
safe_ptr<map,shared>: 35.6 0.45
safe_ptr<map,contfree>: 6.27 2.55
safe<map,contf>rowlock: 5.03 3.18
safe part<mutex>: 3.83 4.18
safe part<contfree>: 1.98 8.08
60 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 5.09 3.14
std::map & std::mutex: 15.1 1.06
safe_ptr<map,mutex>: 15 1.07
safe_ptr<map,shared>: 33.9 0.472
safe_ptr<map,contfree>: 7.97 2.01
safe<map,contf>rowlock: 6.53 2.45
safe part<mutex>: 3.97 4.03
safe part<contfree>: 2.17 7.36
75 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 4.66 3.43
std::map & std::mutex: 14.9 1.07
safe_ptr<map,mutex>: 14.8 1.08
safe_ptr<map,shared>: 30 0.533
safe_ptr<map,contfree>: 8.79 1.82
safe<map,contf>rowlock: 7.6 2.1
safe part<mutex>: 4.11 3.89
safe part<contfree>: 2.36 6.79
90 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 4.77 3.36
std::map & std::mutex: 15.2 1.05
safe_ptr<map,mutex>: 15.4 1.04
safe_ptr<map,shared>: 23.5 0.68
safe_ptr<map,contfree>: 9.7 1.65
safe<map,contf>rowlock: 8.96 1.79
safe part<mutex>: 4.33 3.7
safe part<contfree>: 2.68 5.97
end
After the change:
CPU Cores: 8
Benchmark thread-safe associative containers with size = 100000
Threads = 8, iterations per thread = 2000000
Time & MOps - steady_clock is_steady = 1, num/den = 1 / 1000000000
Latency - high_resolution_clock is_steady = 0, num/den = 1 / 1000000000
burn_cpu() for each multithread operations took: 3.89929 nano-sec
Filling of containers... filled containers.
0 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 9.13 1.75
std::map & std::mutex: 21.6 0.739
safe_ptr<map,mutex>: 21 0.76
safe_ptr<map,shared>: 2.25 7.12
safe_ptr<map,contfree>: 1.51 10.6
safe<map,contf>rowlock: 1.47 10.9
safe part<mutex>: 4.11 3.9
safe part<contfree>: 1.72 9.28
15 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 5.28 3.03
std::map & std::mutex: 15.4 1.04
safe_ptr<map,mutex>: 18.5 0.867
safe_ptr<map,shared>: 28.8 0.556
safe_ptr<map,contfree>: 3.95 4.05
safe<map,contf>rowlock: 2.74 5.85
safe part<mutex>: 3.41 4.7
safe part<contfree>: 1.76 9.07
30 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 4.45 3.59
std::map & std::mutex: 15.5 1.03
safe_ptr<map,mutex>: 15.3 1.05
safe_ptr<map,shared>: 34.7 0.461
safe_ptr<map,contfree>: 5.13 3.12
safe<map,contf>rowlock: 4.07 3.94
safe part<mutex>: 3.85 4.16
safe part<contfree>: 1.83 8.76
45 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 4.67 3.43
std::map & std::mutex: 15.4 1.04
safe_ptr<map,mutex>: 15 1.07
safe_ptr<map,shared>: 35.3 0.454
safe_ptr<map,contfree>: 6.41 2.5
safe<map,contf>rowlock: 5.25 3.05
safe part<mutex>: 4.01 3.99
safe part<contfree>: 2.03 7.88
60 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
std::map - 1 thread: 4.83 3.32
std::map & std::mutex: 15.6 1.03
safe_ptr<map,mutex>: 15.3 1.04
safe_ptr<map,shared>: 32.7 0.489
safe_ptr<map,contfree>: 7.74 2.07
safe<map,contf>rowlock: 6.36 2.52
safe part<mutex>: 4.04 3.96
safe part<contfree>: 2.69 5.94
75 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 4.89 3.27
std::map & std::mutex: 15.3 1.05
safe_ptr<map,mutex>: 15.4 1.04
safe_ptr<map,shared>: 29.8 0.538
safe_ptr<map,contfree>: 8.79 1.82
safe<map,contf>rowlock: 7.65 2.09
safe part<mutex>: 4.26 3.75
safe part<contfree>: 2.45 6.54
90 % of write operations (1/3 insert, 1/3 delete, 1/3 update)
(1 Operation latency, usec)
time, sec MOps Median Min Max
std::map - 1 thread: 4.73 3.38
std::map & std::mutex: 15.7 1.02
safe_ptr<map,mutex>: 15.9 1.01
safe_ptr<map,shared>: 23.7 0.674
safe_ptr<map,contfree>: 9.5 1.68
safe<map,contf>rowlock: 8.77 1.83
safe part<mutex>: 4.48 3.58
safe part<contfree>: 2.64 6.07
end
