[jira] [Reopened] (TAP5-2333) Decrease number of ThreadLocal.get calls

["Michael Mikhulya (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/TAP5-2333?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Michael Mikhulya reopened TAP5-2333:
------------------------------------

Sorry for reopen.
But ~25% performance boost is important.

Is it possible to provide configuration option, so it would be a user choice between operation traces and performance?

My first patch do almost excactly it (operation traces are available in debug mode). We need only introduce configuration option and properly document it.

Btw, it would be great to note following JVM option {{-XX:-OmitStackTraceInFastThrow}}.

> Decrease number of ThreadLocal.get calls
> ----------------------------------------
>
>                 Key: TAP5-2333
>                 URL: https://issues.apache.org/jira/browse/TAP5-2333
>             Project: Tapestry 5
>          Issue Type: Improvement
>            Reporter: Michael Mikhulya
>            Assignee: Howard M. Lewis Ship
>              Labels: performance
>             Fix For: 5.4
>
>         Attachments: 0001-TAP5-2333-Decrease-number-of-ThreadLocal.get-calls.patch, 0002-TAP5-2333-Decrease-number-of-ThreadLocal.get-calls.patch, 0003-TAP5-2333-Decrease-number-of-ThreadLocal.get-calls.patch
>
>
> During profiling I found that ThreadLocal.get is a very hot method call.
> Most frequently it is called from PerThreadOperationTracker.
> PerThreadOperationTracker can be replaced with SimpleOperationTracker which I introduced in a patch.
> SimpleOperationTracker only prints exception without "operations trace".
> "Operations trace" can be useful during debug. So in my patch PerThreadOperationTracker is used in debug mode, but otherwise SimpleOperationTracker is used.
> Please check whether this decision is good for most cases.
> Performance gains are very serious. 
> Time per request decreased on 11ms (23% of overall time).
> All measurements are done with apache benchmark after warm up phase.
> Currently my patch breaks two tests:
> CoreBehaviorsTests. event_handler_return_types
> MiscTests. operation_tracking_via_annotation
> Both tests are broken because tests depend on 'Operation description' which is ignored by SimpleOperationTracker.
> The simplest way to fix tests is to enforce using PerThreadOperationTracker for these tests.
> I'm not sure whether 'Operation description' is definitely useful especially taking into account that only 2 tests become broken. We use SimpleOperationTracker on production for several months already and nobody notice that 'Operation descriptions' are absent. In all cases (for us) it was enough to see a stack trace of exception in logs.
> See TAP5-2332. There is a huge amount of String.format required to track such 'operation descriptions'. By removing calculation of such 'operation descriptions' Tapestry can be made much faster.
> If 'Operation descriptions' is required for some cases than we can introduce some option to enable/disable this feature.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)