RFR: 8274112: (fc) Tune FileChannel.transferTo()

[Brian Burkhalter]

On Wed, 22 Sep 2021 01:25:37 GMT, Brian Burkhalter <bpb at openjdk.org> wrote:

> Please consider this patch which would improve the performance of
> `FileChannel.transferTo()` on macOS and Windows.

On Linux, `sendfile()` was measured to provide faster transfer than either the
trusted channel (mapped buffer copy) or the  arbitrary channel (Java read-write)
paths, and an experimental native read-write path. Therefor on Linux no change is
needed.

On macOS, native read-write was measured to be faster than both the trusted and
arbitrary paths. Benchmark results for various lengths are for the trusted path


Benchmark                    (length)   Mode  Cnt       Score      Error  Units
TransferTune.transferToFC        1000  thrpt   10  127670.489 ± 1018.466  ops/s
TransferTune.transferToFC       10000  thrpt   10  103345.306 ± 1049.713  ops/s
TransferTune.transferToFC      100000  thrpt   10   36488.102 ±  242.382  ops/s
TransferTune.transferToFC      500000  thrpt   10    9440.104 ±   56.543  ops/s
TransferTune.transferToFC     1000000  thrpt   10    4813.599 ±   78.812  ops/s
TransferTune.transferToFC    10000000  thrpt   10     184.298 ±    4.617  ops/s
TransferTune.transferToFC   100000000  thrpt   10      23.391 ±    0.183  ops/s
TransferTune.transferToFC   500000000  thrpt   10       4.674 ±    0.028  ops/s
TransferTune.transferToFC  1000000000  thrpt   10       2.349 ±    0.010  ops/s
TransferTune.transferToFC  2000000000  thrpt   10       0.993 ±    0.004  ops/s
TransferTune.transferToFC  4000000000  thrpt   10       0.455 ±    0.002  ops/s


and the proposed native read-write path


Benchmark                    (length)   Mode  Cnt       Score      Error  Units
TransferTune.transferToFC        1000  thrpt   10  253841.238 ± 1884.480  ops/s
TransferTune.transferToFC       10000  thrpt   10  215629.254 ± 1987.146  ops/s
TransferTune.transferToFC      100000  thrpt   10   48655.924 ± 1060.983  ops/s
TransferTune.transferToFC      500000  thrpt   10   11596.874 ±  126.086  ops/s
TransferTune.transferToFC     1000000  thrpt   10    5889.314 ±   20.170  ops/s
TransferTune.transferToFC    10000000  thrpt   10     364.018 ±    2.149  ops/s
TransferTune.transferToFC   100000000  thrpt   10      31.956 ±    0.277  ops/s
TransferTune.transferToFC   500000000  thrpt   10       6.282 ±    0.090  ops/s
TransferTune.transferToFC  1000000000  thrpt   10       3.134 ±    0.103  ops/s
TransferTune.transferToFC  2000000000  thrpt   10       1.542 ±    0.080  ops/s
TransferTune.transferToFC  4000000000  thrpt   10       0.750 ±    0.087  ops/s


On Windows the native read-write path was also measured to be  faster than the trusted path except that beyond a certain limit the trusted path becomes faster than native read-write. The value of this limit appears to be approximately 2MB but more verification might be needed.

-------------

PR: https://git.openjdk.java.net/jdk/pull/5623