Hi all! I recently submitted 8 enhancement tickets via Bugzilla, all related to JSP performance, and I thought an explanation would be helpful.
In early September, I received an internal report that a small EL change created a 40ms latency regression on the affected calls. The original EL expression was already complex, and the change added a relatively simple clause at the end of a series of "&&" . On the face of it, the data wasn't rational, and I was deeply skeptical. However, research into the execution of that statement identified a number of Tomcat hotspots, many of which are fixable with isolated changes - thus most of the tickets. One or two are less isolated and deserve more caution. I asked the change implementer to simplify the EL statement by moving much of the logic into Java; this reached production on Monday and we confirmed this eliminated the 40ms regression. I can't adequately explain how this change reached 40ms of impact, but my best theory is that our allocation-heavy application was pushed over a threshold by the EL's burst of extra allocation. That threshold may be TLAB, or contention on a JVM-internal lock, or something else I haven't thought of. I welcome any other possible explanations. Relevant tickets are: https://bz.apache.org/bugzilla/show_bug.cgi?id=69338 - parsing optimization when turning EL -> AST nodes (resolved) https://bz.apache.org/bugzilla/show_bug.cgi?id=69348 - allocation reduction in ELParser (resolved) https://bz.apache.org/bugzilla/show_bug.cgi?id=69377 - JSP work reduction, large scope (pending) https://bz.apache.org/bugzilla/show_bug.cgi?id=69381 - allocation reduction when executing reflecting expressions, i.e. "bean.myMethod()" (PR submitted) https://bz.apache.org/bugzilla/show_bug.cgi?id=69382 - performance optimization to generated code for <jsp:param> (patch submitted) https://bz.apache.org/bugzilla/show_bug.cgi?id=69398 - allocation reduction in BodyContentImpl (pending) https://bz.apache.org/bugzilla/show_bug.cgi?id=69406 - bug fix that enables existing enum optimizations in one of our applications (pending) https://bz.apache.org/bugzilla/show_bug.cgi?id=69419 - performance optimization to ApplicationHttpRequest (patch submitted) The most impactful of the pending tickets is #69381 (reflective expressions), which removes object allocation that scales per method on a bean. I plan to add patches or PRs to the remaining issues in the next couple of weeks. Thanks everyone for their help and attention - and again, I welcome any theories about how the change reached 40ms of impact. John
