C++ benchmark for default, pmr and custom local allocator
These test are fast and dirty, only to see if there's any potential benefit in what I usually do coding. Do not exepect code of any quality.
Results of 1000 iterations for std::list<int> of size 250000
w (warming step) : 4.235 sec
t1 (default std alloc no reserve) : 3.423 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 4.093 sec; t1/t2: 0.836
t3 (pmr alloc no buffer no reserve) : 2.018 sec; t1/t3: 1.696
t4 (pmr alloc with buffer with reserve) : 1.603 sec; t1/t4: 2.135
t5 (custom stack alloc with reserve) : 0.550 sec; t1/t5: 6.224
t6 (custom stackalloc allctd once, resvd): 0.604 sec; t1/t5: 5.667
Results of 268435456 iterations for std::vector<int> of size 4
w (warming step) : 17.327 sec
t0 (default std alloc with reserve) : 7.593 sec; t1/t0: 1.709
t1 (default std alloc no reserve) : 12.977 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 15.632 sec; t1/t2: 0.830
t3 (pmr alloc no buffer no reserve) : 15.754 sec; t1/t3: 0.824
t4 (pmr alloc with buffer with reserve) : 7.680 sec; t1/t4: 1.690
t5 (custom stackalloc with reserve) : 4.400 sec; t1/t5: 2.949
t6 (custom stackalloc allctd once, resvd): 4.293 sec; t1/t5: 3.023
Results of 67108864 iterations for std::vector<int> of size 16
w (warming step) : 7.234 sec
t0 (default std alloc with reserve) : 2.429 sec; t1/t0: 2.102
t1 (default std alloc no reserve) : 5.107 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 6.301 sec; t1/t2: 0.811
t3 (pmr alloc no buffer no reserve) : 5.254 sec; t1/t3: 0.972
t4 (pmr alloc with buffer with reserve) : 2.410 sec; t1/t4: 2.119
t5 (custom stackalloc with reserve) : 2.134 sec; t1/t5: 2.393
t6 (custom stackalloc allctd once, resvd): 2.094 sec; t1/t5: 2.439
Results of 16777216 iterations for std::vector<int> of size 64
w (warming step) : 3.387 sec
t0 (default std alloc with reserve) : 1.197 sec; t1/t0: 1.834
t1 (default std alloc no reserve) : 2.196 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 2.685 sec; t1/t2: 0.818
t3 (pmr alloc no buffer no reserve) : 2.133 sec; t1/t3: 1.029
t4 (pmr alloc with buffer with reserve) : 1.238 sec; t1/t4: 1.773
t5 (custom stackalloc with reserve) : 1.190 sec; t1/t5: 1.845
t6 (custom stackalloc allctd once, resvd): 1.193 sec; t1/t5: 1.840
Results of 4194304 iterations for std::vector<int> of size 256
w (warming step) : 2.196 sec
t0 (default std alloc with reserve) : 0.840 sec; t1/t0: 1.517
t1 (default std alloc no reserve) : 1.275 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 1.426 sec; t1/t2: 0.894
t3 (pmr alloc no buffer no reserve) : 1.329 sec; t1/t3: 0.959
t4 (pmr alloc with buffer with reserve) : 0.916 sec; t1/t4: 1.391
t5 (custom stackalloc with reserve) : 0.924 sec; t1/t5: 1.379
t6 (custom stackalloc allctd once, resvd): 0.920 sec; t1/t5: 1.386
Results of 1048576 iterations for std::vector<int> of size 1024
w (warming step) : 1.852 sec
t0 (default std alloc with reserve) : 0.753 sec; t1/t0: 1.323
t1 (default std alloc no reserve) : 0.996 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 1.047 sec; t1/t2: 0.951
t3 (pmr alloc no buffer no reserve) : 1.031 sec; t1/t3: 0.965
t4 (pmr alloc with buffer with reserve) : 0.835 sec; t1/t4: 1.192
t5 (custom stackalloc with reserve) : 0.857 sec; t1/t5: 1.162
t6 (custom stackalloc allctd once, resvd): 0.853 sec; t1/t5: 1.167
Results of 262144 iterations for std::vector<int> of size 4096
w (warming step) : 2.309 sec
t0 (default std alloc with reserve) : 0.727 sec; t1/t0: 2.011
t1 (default std alloc no reserve) : 1.462 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 0.920 sec; t1/t2: 1.589
t3 (pmr alloc no buffer no reserve) : 0.908 sec; t1/t3: 1.610
t4 (pmr alloc with buffer with reserve) : 0.824 sec; t1/t4: 1.775
t5 (custom stackalloc with reserve) : 0.847 sec; t1/t5: 1.725
t6 (custom stackalloc allctd once, resvd): 0.844 sec; t1/t5: 1.733
Results of 65536 iterations for std::vector<int> of size 16384
w (warming step) : 2.599 sec
t0 (default std alloc with reserve) : 0.731 sec; t1/t0: 2.410
t1 (default std alloc no reserve) : 1.761 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 0.892 sec; t1/t2: 1.974
t3 (pmr alloc no buffer no reserve) : 0.887 sec; t1/t3: 1.985
t4 (pmr alloc with buffer with reserve) : 0.812 sec; t1/t4: 2.170
t5 (custom stackalloc with reserve) : 0.838 sec; t1/t5: 2.102
t6 (custom stackalloc allctd once, resvd): 0.835 sec; t1/t5: 2.108
Results of 16384 iterations for std::vector<int> of size 65536
w (warming step) : 2.656 sec
t0 (default std alloc with reserve) : 0.726 sec; t1/t0: 2.500
t1 (default std alloc no reserve) : 1.816 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 0.886 sec; t1/t2: 2.050
t3 (pmr alloc no buffer no reserve) : 0.881 sec; t1/t3: 2.060
t4 (pmr alloc with buffer with reserve) : 0.814 sec; t1/t4: 2.231
t5 (custom stackalloc with reserve) : 0.839 sec; t1/t5: 2.165
t6 (custom stackalloc allctd once, resvd): 0.835 sec; t1/t5: 2.173
Results of 4096 iterations for std::vector<int> of size 262144
w (warming step) : 2.671 sec
t0 (default std alloc with reserve) : 0.728 sec; t1/t0: 2.517
t1 (default std alloc no reserve) : 1.832 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 0.904 sec; t1/t2: 2.028
t3 (pmr alloc no buffer no reserve) : 0.901 sec; t1/t3: 2.034
t4 (pmr alloc with buffer with reserve) : 0.815 sec; t1/t4: 2.248
t5 (custom stackalloc with reserve) : 0.841 sec; t1/t5: 2.178
t6 (custom stackalloc allctd once, resvd): 0.837 sec; t1/t5: 2.188
Results of 1024 iterations for std::vector<int> of size 1048576
w (warming step) : 3.791 sec
t0 (default std alloc with reserve) : 0.729 sec; t1/t0: 4.046
t1 (default std alloc no reserve) : 2.950 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 1.944 sec; t1/t2: 1.518
t3 (pmr alloc no buffer no reserve) : 1.927 sec; t1/t3: 1.531
t4 (pmr alloc with buffer with reserve) : 0.815 sec; t1/t4: 3.619
t5 (custom stackalloc with reserve) : 0.841 sec; t1/t5: 3.509
t6 (custom stackalloc allctd once, resvd): 0.838 sec; t1/t5: 3.521
Results of 1000000 iterations for std::vector<int> of max size 256
w (warming step) : 0.326 sec
t0 (default std alloc with reserve) : 0.136 sec; t1/t0: 1.499
t1 (default std alloc no reserve) : 0.203 sec; t1/t1: 1.000
t2 (default pmr alloc no reserve) : 0.241 sec; t1/t2: 0.843
t3 (pmr alloc no buffer no reserve) : 0.216 sec; t1/t3: 0.940
t4 (pmr alloc with buffer with reserve) : 0.152 sec; t1/t4: 1.335
t7 (custom stackalloc with backup, resvd): 0.150 sec; t1/t5: 1.359
(in this test the final size is not known, "reserves" are made with
a guess, in average equal to half to the final size)
The simple stack allocator (which, in the best case scenario, when finishes its buffer causes a segfault), is quite efficient when compared to all other general or local allocators. This would be very interesting if the simple "reserve" wouldn't be almost always better (aside for very small vector sizes). Notice also that if we use this version of stack_allocator, we must know the max size beforehand, so we can call reserve with vector using the same size. For the list case (which, as far as I know, has no reserve method (as one would expect)), it can be a useful way to implement a "reserve" functionality, but only with the general stack_allocator that call a general allocator when the buffer finishes.
Comparing t5 and t6, allocating a stack buffer in the parent function seems to be effective with very small sizes. It rapidly loosees any relevance (as the stack-allocator compared to the simple "reserve" call of std::vector).
Finally, when the final size is not known, making a guess with "reserve" still seems more efficient than allocating into the stack a small buffer for smaller sizes.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent