RFR: 8268435: (ch) ChannelInputStream could override readAllBytes

[Brian Burkhalter]

On Thu, 23 Sep 2021 01:52:31 GMT, Brian Burkhalter <bpb at openjdk.org> wrote:

> This change would override `readAllBytes()` and `readNBytes(int)` in `ChannelInputStream` thereby improving performance for all but smaller streams.

Throughput measured without the change for various stream lengths was

Benchmark                          (length)   Mode  Cnt       Score      Error  Units
ChannelInputStream.readAllBytes        1000  thrpt   10  418393.824 ± 2969.676  ops/s
ChannelInputStream.readAllBytes       10000  thrpt   10  196087.387 ± 1947.272  ops/s
ChannelInputStream.readAllBytes      100000  thrpt   10   29089.090 ±  430.841  ops/s
ChannelInputStream.readAllBytes     1000000  thrpt   10    2534.655 ±    5.373  ops/s
ChannelInputStream.readAllBytes    10000000  thrpt   10     209.193 ±    1.106  ops/s
ChannelInputStream.readAllBytes   100000000  thrpt   10      20.010 ±    0.047  ops/s
ChannelInputStream.readAllBytes  1000000000  thrpt   10       1.455 ±    0.204  ops/s
ChannelInputStream.readNBytes          1000  thrpt   10  544002.863 ± 1900.158  ops/s
ChannelInputStream.readNBytes         10000  thrpt   10  421766.791 ± 2928.367  ops/s
ChannelInputStream.readNBytes        100000  thrpt   10   57035.078 ±  214.834  ops/s
ChannelInputStream.readNBytes       1000000  thrpt   10    5855.087 ±   52.283  ops/s
ChannelInputStream.readNBytes      10000000  thrpt   10     427.560 ±    2.925  ops/s
ChannelInputStream.readNBytes     100000000  thrpt   10      40.118 ±    0.091  ops/s
ChannelInputStream.readNBytes    1000000000  thrpt   10       3.315 ±    0.434  ops/s

and with the change in place was

Benchmark                          (length)   Mode  Cnt       Score      Error  Units
ChannelInputStream.readAllBytes        1000  thrpt   10  337785.548 ± 2702.517  ops/s
ChannelInputStream.readAllBytes       10000  thrpt   10  245969.721 ± 1362.465  ops/s
ChannelInputStream.readAllBytes      100000  thrpt   10   75736.303 ±  321.443  ops/s
ChannelInputStream.readAllBytes     1000000  thrpt   10    8835.464 ±   73.547  ops/s
ChannelInputStream.readAllBytes    10000000  thrpt   10     370.141 ±    0.675  ops/s
ChannelInputStream.readAllBytes   100000000  thrpt   10      34.029 ±    0.152  ops/s
ChannelInputStream.readAllBytes  1000000000  thrpt   10       3.414 ±    0.032  ops/s
ChannelInputStream.readNBytes          1000  thrpt   10  354884.272 ± 2264.808  ops/s
ChannelInputStream.readNBytes         10000  thrpt   10  296775.457 ± 4296.873  ops/s
ChannelInputStream.readNBytes        100000  thrpt   10  120579.221 ±  374.626  ops/s
ChannelInputStream.readNBytes       1000000  thrpt   10   18079.568 ±  109.945  ops/s
ChannelInputStream.readNBytes      10000000  thrpt   10     860.772 ±    2.333  ops/s
ChannelInputStream.readNBytes     100000000  thrpt   10      69.266 ±    0.293  ops/s
ChannelInputStream.readNBytes    1000000000  thrpt   10       6.840 ±    0.016  ops/s

In effect, degraded performance for very small lengths is traded off against better performance for larger lengths. Note however that the `readNBytes()` measurements were made for `len = length/2` and so are less telling. Otherwise said, the `readAllBytes()` results are more straightforwardly interpreted. Adding a threshold was not investigated.

Unfortunately this is mostly code duplication from `FileInputStream`. It would be better if there were a common location where the implementation could be placed and used by both cases.

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

PR: https://git.openjdk.java.net/jdk/pull/5645