<div dir="ltr">Hey!<div><br></div><div>Please have a look at the latest perf test results [1][2]:</div><div><div><br></div><div><div>Benchmark Mode Cnt Score Error Units</div><div>MyBenchmark.noVariable thrpt 20 <b>681131269.875</b> ± 3961932.923 ops/s</div><div>MyBenchmark.withIsTraceEnabledCheck thrpt 20 <b>676307984.921</b> ± 14305970.393 ops/s</div><div>MyBenchmark.withVariable thrpt 20 <b>2411000894.582</b> ± 17382438.635 ops/s</div></div></div><div><br></div><div>I think there is no surprise here.. using a field, which stores the result of `logger.isTraceEnabled()` evaluation is 3 times faster than other options.</div><div><br></div><div>If anyone is interested in printing out JIT stuff, I also ran it with "-XX:+PrintCompilation", "-XX:+PrintCompilation2" and "-XX:+PrintInlining" here [3]. </div><div><br></div><div>I'm not a performance expert but it seems that JIT could not inline the "translate" method because of its size (see line 1861). However it tried several times with different optimizations (and some of them were thrown away - "made not entrant" messages [4]).</div><div><br></div><div>Let's wait for James' opinion on this, but I think we should address this issue on JBoss Logging/LogManager side (so I agree with David here) and make those parts inlinable (wow, does this word even exist? :D). Once this is done, we could experiment further in Infinispan codebase and see how this relates to some real world benchmarks...</div><div><br></div><div>Thanks</div><div>Sebastian</div><div><br></div><div>[1] <a href="https://gist.github.com/slaskawi/6766b6e17c7a28ac8d8962293c48a53c">https://gist.github.com/slaskawi/6766b6e17c7a28ac8d8962293c48a53c</a></div><div>[2] Test repository: <a href="https://github.com/slaskawi/jboss-logging-perf-test">https://github.com/slaskawi/jboss-logging-perf-test</a></div><div>[3] <a href="https://gist.github.com/slaskawi/6f317bb05539611434bc91d66924bae0">https://gist.github.com/slaskawi/6f317bb05539611434bc91d66924bae0</a></div><div>[4] <a href="https://www.safaribooksonline.com/library/view/java-performance-the/9781449363512/ch04.html">https://www.safaribooksonline.com/library/view/java-performance-the/9781449363512/ch04.html</a></div><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Oct 3, 2016 at 7:32 AM, Sebastian Laskawiec <span dir="ltr"><<a href="mailto:slaskawi@redhat.com" target="_blank">slaskawi@redhat.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr">Hey!<div><br></div><div>Adding James to this thread.</div><div><br></div><div>@Dennis - I think Dan has a point here. The trick with checking a field in a class is 3 times faster. Most of the checks are done in core so they are executed multiple times per operation. Changing all those places is probably not an option.</div><div><br></div><div>@David - Let me run a test with JBoss Log Manager and get back to you with some results. But if Dan is right, and the problem is with enum mapping, I will get similar results.</div><div><br></div><div>Thanks</div><span class="gmail-HOEnZb"><font color="#888888"><div>Sebastian</div></font></span></div><div class="gmail-HOEnZb"><div class="gmail-h5"><div class="gmail_extra"><br><div class="gmail_quote">On Sat, Oct 1, 2016 at 11:53 AM, Dan Berindei <span dir="ltr"><<a href="mailto:dan.berindei@gmail.com" target="_blank">dan.berindei@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><span>On Fri, Sep 30, 2016 at 10:15 PM, David M. Lloyd <<a href="mailto:david.lloyd@redhat.com" target="_blank">david.lloyd@redhat.com</a>> wrote:<br>
> The performance problem that this trick is meant to resolve is really a<br>
> problem in the logging backend. It *should* be faster inside of<br>
> WildFly, where JBoss LogManager is used, because that project just<br>
> checks a single volatile field for the level check... and the path to<br>
> that code *should* be inline-friendly.<br>
><br>
<br>
</span>Indeed, we started using this trick because of log4j 1.2, which needs<br>
to walk the logger hierarchy in order to check the level, and it made<br>
a significant difference there.<br>
<br>
Nowadays I think it's pretty close to optimal in all logging<br>
frameworks. The only nitpick is that they all use enums for the<br>
levels, and the JIT can't inline Level.TRACE.value as it would with a<br>
Level.TRACE_VALUE int constant. If JDK9 fixes that, then it's going to<br>
be more or less equivalent to using a volatile "trace" field in each<br>
class, so it should be suitable even for local read operations that<br>
take < 200ns.<br>
<br>
We'd probably still need to weed out some of the trace messages, as we<br>
probably have more than 10 of them during such a read operation. I<br>
confess that I added way too many trace logs myself, precisely because<br>
I knew we are using a static final field and the JIT compiler doesn't<br>
even generate code for that branch.<br>
<br>
Cheers<br>
<span><font color="#888888">Dan<br>
</font></span><div><div>______________________________<wbr>_________________<br>
infinispan-dev mailing list<br>
<a href="mailto:infinispan-dev@lists.jboss.org" target="_blank">infinispan-dev@lists.jboss.org</a><br>
<a href="https://lists.jboss.org/mailman/listinfo/infinispan-dev" rel="noreferrer" target="_blank">https://lists.jboss.org/mailma<wbr>n/listinfo/infinispan-dev</a><br>
</div></div></blockquote></div><br></div>
</div></div></blockquote></div><br></div></div>