Feature #20902
closed
Allow `IO::Buffer#copy` to release the GVL.
Description
Related to https://bugs.ruby-lang.org/issues/20876.
Background¶
IO::Buffer#copy execution time is proportional to the length of the data copied. As such, large copies can take a long time (100ms+). Currently, the GVL is not released, which can stall the Ruby interpreter.
Proposal¶
Pull Request: https://github.com/ruby/ruby/pull/12021
If the size of the data to be copied is larger than a specific amount (heuristic), we will perform memmove using rb_nogvl.
The initial size heuristic is set to 1MiB. This won't be perfect for every system, but should be good enough to avoid ms+ stalls.
Results¶
I measured the difference:
| GVL | Threads | Buffer Size | Total Duration | Throughput (MB/s) |
|---|---|---|---|---|
| Yes | 1 | 1 | 0.12ms | 8393.09 |
| Yes | 1 | 5 | 0.51ms | 9857.7 |
| Yes | 1 | 10 | 1.12ms | 8937.54 |
| Yes | 1 | 20 | 2.22ms | 9015.95 |
| Yes | 2 | 1 | 0.24ms | 8307.07 |
| Yes | 2 | 5 | 1.13ms | 8819.58 |
| Yes | 2 | 10 | 1.49ms | 13385.35 |
| Yes | 2 | 20 | 5.63ms | 7110.8 |
| Yes | 4 | 1 | 0.92ms | 4360.18 |
| Yes | 4 | 5 | 2.08ms | 9606.58 |
| Yes | 4 | 10 | 4.51ms | 8863.13 |
| Yes | 4 | 20 | 9.3ms | 8601.41 |
| Yes | 8 | 1 | 1.22ms | 6574.93 |
| Yes | 8 | 5 | 3.56ms | 11239.27 |
| Yes | 8 | 10 | 7.31ms | 10943.68 |
| Yes | 8 | 20 | 15.57ms | 10274.99 |
| Yes | 16 | 1 | 1.95ms | 8220.16 |
| Yes | 16 | 5 | 5.51ms | 14518.05 |
| Yes | 16 | 10 | 13.77ms | 11618.96 |
| Yes | 16 | 20 | 27.21ms | 11759.43 |
| Yes | 32 | 1 | 3.24ms | 9891.05 |
| Yes | 32 | 5 | 11.42ms | 14007.41 |
| Yes | 32 | 10 | 21.64ms | 14786.48 |
| Yes | 32 | 20 | 45.52ms | 14060.25 |
| No | 1 | 1 | 0.13ms | 7582.85 |
| No | 1 | 5 | 0.44ms | 11248.55 |
| No | 1 | 10 | 1.11ms | 9029.91 |
| No | 1 | 20 | 2.43ms | 8228.42 |
| No | 2 | 1 | 0.18ms | 11245.61 |
| No | 2 | 5 | 0.96ms | 10396.76 |
| No | 2 | 10 | 1.9ms | 10501.59 |
| No | 2 | 20 | 3.16ms | 12656.77 |
| No | 4 | 1 | 0.69ms | 5827.76 |
| No | 4 | 5 | 1.15ms | 17440.54 |
| No | 4 | 10 | 2.31ms | 17307.79 |
| No | 4 | 20 | 4.11ms | 19483.68 |
| No | 8 | 1 | 0.67ms | 11954.1 |
| No | 8 | 5 | 1.3ms | 30713.68 |
| No | 8 | 10 | 2.05ms | 38990.98 |
| No | 8 | 20 | 4.15ms | 38552.37 |
| No | 16 | 1 | 0.96ms | 16698.03 |
| No | 16 | 5 | 1.46ms | 54782.47 |
| No | 16 | 10 | 2.74ms | 58295.64 |
| No | 16 | 20 | 4.89ms | 65482.43 |
| No | 32 | 1 | 1.82ms | 17554.27 |
| No | 32 | 5 | 2.68ms | 59673.59 |
| No | 32 | 10 | 3.87ms | 82733.34 |
| No | 32 | 20 | 6.93ms | 92297.47 |
In the base case, the performance is about the same, but in the best case, the throughput is significantly better: 15GiB/s vs 92GiB/s (32 threads copying 20MiB of data).
Updated by 
Updated by Anonymous