On Wed, 16 Feb 2022 09:30:46 GMT, Volker Simonis <simonis at openjdk.org> wrote:
> Currently, `InflaterInputStream::read()` first does a native call to the underlying zlib `inflate()` function and only afterwards checks if the inflater requires input (i.e. `Inflater::needsInput()`) or has finished inflating (`Inflater::finished()`). This leads to an unnecessary native call to `inflate()` when `InflaterInputStream::read()` is invoked for the very first time because `Inflater::fill()` hasn't been called and another unnecessary native call to detect the end of the stream. For small streams/files which completely fit into the output buffer passed to `InflaterInputStream::read()` we can therefore save two out of three native calls. The attached micro benchmark shows that this results in a 5%-10% performance improvement for zip files of sizes between 4096 to 256 bytes (notice that in common jars like Tomcat, Spring-Boot, Maven, Jackson, etc. about 60-80% of the classes are smaller than 4096 bytes).
>
>
> before JDK-8281962
> ------------------
> Benchmark (size) Mode Cnt Score Error Units
> InflaterInputStreams.inflaterInputStreamRead 256 avgt 5 2.571 ± 0.120 us/op
> InflaterInputStreams.inflaterInputStreamRead 512 avgt 5 2.861 ± 0.064 us/op
> InflaterInputStreams.inflaterInputStreamRead 4096 avgt 5 5.110 ± 0.278 us/op
>
> after JDK-8281962
> -----------------
> Benchmark (size) Mode Cnt Score Error Units
> InflaterInputStreams.inflaterInputStreamRead 256 avgt 5 2.332 ± 0.081 us/op
> InflaterInputStreams.inflaterInputStreamRead 512 avgt 5 2.691 ± 0.293 us/op
> InflaterInputStreams.inflaterInputStreamRead 4096 avgt 5 4.812 ± 1.038 us/op
>
>
> Tested with the JTreg zip/jar/zipfs and the JCK zip/jar tests.
>
> As a side effect, this change also fixes an issue with alternative zlib implementations like zlib-Chromium or zlib-Cloudflare which pad the inflated bytes with a specif byte pattern at the end of the output for debugging purpose. This breaks code patterns like the following:
>
>
> int readcount = 0;
> while ((bytesRead = inflaterInputStream.read(data, 0, bufferSize)) != -1) {
> outputStream.write(data, 0, bytesRead);
> readCount++;
> }
> if (readCount == 1) {
> return data; // <---- first bytes might be overwritten
> }
> return outputStream.toByteArray();
>
>
> Even if the whole data fits into the `data` array during the first call to `inflaterInputStream.read()`, we still need a second call to `inflaterInputStream.read()` in order to detect the end of the inflater stream. If this second call calls into the native `inflate()` function of Cloudflare/Chromium's zlib version, this will still write some padding bytes at the beginning of the `data` array and overwrite the previously read data. This issue has been reported in Spring [1] and ASM [2] when using these libraries with Cloudflares zlib version (by setting `LD_LIBRARY_PATH` accordingly).
>
> [1] https://github.com/spring-projects/spring-framework/issues/27429
> [2] https://gitlab.ow2.org/asm/asm/-/issues/317955
Code changes look good to me.
(Shouldn't the copyright header have a year?)
test/micro/org/openjdk/bench/java/util/zip/InflaterInputStreams.java line 78:
> 76: private byte[] words;
> 77: private static final int MAX_SIZE = 5000; // Should be bigger than the biggest size @Param
> 78: private static byte[] inflated = new byte[MAX_SIZE];
If it isn't important for the benchmark, I'd remove the hard-coded `MAX_SIZE` and make `inflated` non-static and allocate it in `beforeRun` using `size` (or some multiple thereof). Future enhancements could very well be interested in assessing the performance of larger sized entries, and we shouldn't put up trip-wires if they want to increase `size`.
-------------
Changes requested by redestad (Reviewer).
PR: https://git.openjdk.java.net/jdk/pull/7492