[ 
https://issues.apache.org/jira/browse/TS-159?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Leif Hedstrom updated TS-159:
-----------------------------

    Attachment: TS-159.diff

This diff fixes the problem, ignore the "indentation" fixes (indent did a 
crappy job on this file, I fixed the worst of it). The issue is the 
initialization part of the ctor, I think this gets called when iObject is 
subclassed, which is wrong becuse the size is already properly set in the 
::operator new.

I'm not a C++ expert in any way, so maybe jplevyak can correct me if this is 
wrong. It does seem to solve the problem though.

> Severe performance degradation with transaction logging enabled.
> ----------------------------------------------------------------
>
>                 Key: TS-159
>                 URL: https://issues.apache.org/jira/browse/TS-159
>             Project: Traffic Server
>          Issue Type: Bug
>          Components: Logging
>            Reporter: Leif Hedstrom
>            Assignee: Leif Hedstrom
>            Priority: Critical
>             Fix For: 2.0.0a
>
>         Attachments: TS-159.diff
>
>
> With logging enabled, e.g.
> CONFIG proxy.config.log2.logging_enabled INT 3
> CONFIG proxy.config.log2.squid_log_enabled INT 1
> CONFIG proxy.config.log2.squid_log_is_ascii INT 0
> ATS "trunk" (and "dev" branch) suffers from serious performance degradation 
> over time. I ran a number of tests, each 30s long, fetching a 500 byte long 
> object from cache (ram cache enabled). The tests starts off doing about 20k 
> QPS on my box, but with every iteration, it degrades (more so in the 
> beginning). Within 30 minutes, performance had dropped to 3,800 QPS, at which 
> point the process segfaulted (I think it's a meaningless trace though, but 
> attaching it nonetheless).
> The CPU is completely pegged at all times, so it's definitely doing less and 
> less workload, using the same (or more) CPU. I did not observe any memory 
> growth, so I don't think it's a memory leak per se (but could still obviously 
> be related to starving some resource inside TS).

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.

Reply via email to