Heavy Allocation in Emitter.analyzeScalar(String) due to Regex Overhead

[Carter Kozak]

Writing large yaml documents using snakeyml I’ve found GC contention resulting from integer array allocations within Pattern.matcher.
I believe we can dramatically improve performance by refactoring the leading-zero check away from regex and into a simple utility method along these lines:

private static boolean hasLeadingZero(String scalar) {
if (scalar.length() > 1 && scalar.charAt(0) == '0') {
int secondChar = scalar.charAt(1);
return (secondChar >= '0' && secondChar <= '9') || secondChar == '_';
}
return false;
}

[Carter Kozak]

I’ve created a PR here with my proposed solution: https://github.com/snakeyaml/snakeyaml/pull/5

There are a few similar regex patterns. I don’t see them in profiling data, so I opted not to attempt replacement. In a fairly large application, integer arrays within analyzeScalar regex account for 83% of total allocation exceeding 9 terabytes per minute.

[Andrey Somov]

The tests fail. Please run the tests locally

./mvn clean verify

[Andrey Somov]

you only check first and second char.

[Andrey Somov]

Can you please contribute the test where you measure the memory consumption ?

[Carter Kozak]

I’ve put together a test here: https://github.com/snakeyaml/snakeyaml/pull/6 since the original PR was merged while I worked on writing the test (thank you for such a quick review!). This sort of test is tricky because garbage collectors are complex machines, and it can be challenging to test specifically what we mean to without making the build unnecessarily expensive or flaky.

[Andrey Somov]

How do we know that anything is improved ?
What was the setup for memory consumption ?

[Carter Kozak]

Here’s the JMH benchmark I described earlier, which does a better job of illustrating the high rate of allocation: https://github.com/snakeyaml/snakeyaml/pull/7

Benchmark results and analysis copied here from the github pr for posterity.

Before 26f5d5f8 (benchmarks run atop 98ff897):

Benchmark                                        Mode  Cnt     Score     Error   Units
EmitterBenchmark.emitScalar                      avgt    3     0.311 ±   0.009   us/op
EmitterBenchmark.emitScalar:·gc.alloc.rate       avgt    3  6697.479 ± 420.474  MB/sec
EmitterBenchmark.emitScalar:·gc.alloc.rate.norm  avgt    3   256.000 ±   0.001    B/op
EmitterBenchmark.emitScalar:·gc.count            avgt    3   108.000            counts
EmitterBenchmark.emitScalar:·gc.time             avgt    3    62.000                ms

After 26f5d5f8 (benchmarks run atop 8a3063b):

Benchmark                                        Mode  Cnt     Score     Error   Units
EmitterBenchmark.emitScalar                      avgt    3     0.220 ±   0.006   us/op
EmitterBenchmark.emitScalar:·gc.alloc.rate       avgt    3  4713.283 ± 303.238  MB/sec
EmitterBenchmark.emitScalar:·gc.alloc.rate.norm  avgt    3   128.000 ±   0.001    B/op
EmitterBenchmark.emitScalar:·gc.count            avgt    3   116.000            counts
EmitterBenchmark.emitScalar:·gc.time             avgt    3    63.000                ms

The benchmarks were run on a macbook, so cpu scaling makes the timing values a bit chaotic and I wouldn't put too much trust in the ~30% speedup shown. The important piece is gc.alloc.rate.norm which is the garbage collected bytes per emitScalar invocation, which cut in half.

[Carter Kozak]

Resolving this because https://bitbucket.org/snakeyaml/snakeyaml/commits/26f5d5f82668aa396ca201b524159ab45158d565 has merged.

[Andrey Somov]

How do you produce the result ? Can you please provide the complete instructions to reproduce it ?

[Carter Kozak]

Using the code change in https://github.com/snakeyaml/snakeyaml/pull/7, I executed this main method in my IDE: https://github.com/snakeyaml/snakeyaml/pull/7/files#diff-0dced93986ac61628c4b3093981408316cc847f3db657af126ed671b92d1f5cdR60-R63

I opted for a simple configuration, using a separate module for benchmarks is more standard, based on this maven archetype: https://github.com/openjdk/jmh#preferred-usage-command-line

[Andrey Somov]

It will be delivered in version 2.1

https://bitbucket.org/snakeyaml/snakeyaml/wiki/Changes