[jira] Commented: (TS-157) Improve stats API

["Raghav Jeyaraman (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/TS-157?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12835947#action_12835947 ] 

Raghav Jeyaraman commented on TS-157:
-------------------------------------

We did a few prototype implementations and measured their performance. 

Hardware used for testing:
Processors:     2 x Xeon L5320 1.86GHz (8 cores)
Memory:         7.8GB

Methodology: 
We wrote a plugin in which the stats were registered in the INKPluginInit function. The plugin had code to increment all stats on INK_HTTP_CACHE_LOOKUP_COMPLETE_HOOK, which essentially means that all the stats were incremented for any request that came in.

Mean response times for each of the tests follow.

hash_map + INKMutex : 
Global hash map for all stats, protected by INKMutex. 
2500 stats, 1000 parallel fetches - 29520.7ms
50 stats, 1000 parallel fetches - 1051.58ms

Per thread map<string, int>:
2500 stats, 1000 parallel fetches - 3666.67ms 

Per thread hash_map<string, int>:
2500 stats, 1000 parallel fetches - 1091.36ms

Per thread vector<int> with a map<int, string>:
On registering a stat, position in the vector at which the stat is held is returned. This is used for subsequent increments on the stat.
A map maintains a mapping from the position of the stat to the stat name which was used for registering it.
2500 stats, 1000 parallel fetches - 420.941ms
225 stats, 1000 parallel fetches - 184.773ms

Current INKStat:
250 stats, 1000 parallel fetches - 420.941ms
225 stats, 1000 parallel fetches - 184.773ms

>From the above numbers, the vector based implementation outperforms the current INKStat implementation.
Collection of stats from all the threads happens in a separate continuation, just like the current implementation.

Thanks,
--raghav




> Improve stats API
> -----------------
>
>                 Key: TS-157
>                 URL: https://issues.apache.org/jira/browse/TS-157
>             Project: Traffic Server
>          Issue Type: Improvement
>          Components: Stats
>            Reporter: Raghav Jeyaraman
>            Assignee: Raghav Jeyaraman
>
> As it stands today, traffic server limits the number of stats that can be registered to 250. 
> The code must be changed to allow a large number of stats (100000)  to be tracked.
> The API for adding coupled stats is cumbersome to use and as a result, has very few users. Current stats API must be cleaned up or gotten rid of.
> Thanks,
> --raghav

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