Re: some profiling results:

From: Andres Kroonmaa <andre@dont-contact.us>
Date: Wed, 19 Feb 2003 20:37:57 +0200

 Robert,

 What are you after? prof shows garbage. Get the cpuProfiling to measure
 what you want. Its no way xprof_* routines take up 2% of time, thats
 ridiculous. They run 10-20 cpu opcodes max.
 prof is also very intrusive into code, changing true cpu usage splits alot.
 You may use it to get idea of call graphs, but actual measurements
 should be done with xprof probes. Also note that things change dramatically
 when cpu becomes a bottleneck and there are thousands of FDs open.

 I'd focus of profiling when cpu usage is at ~100%. Which shows me that
 most problems we have are in comm_poll and acls (my workload).
 To be gained - ~ 15-40%

On 20 Feb 2003, at 0:15, Robert Collins <robertc@squid-cache.org> wrote:

> Flat profile:
>
> Each sample counts as 0.01 seconds.
> % cumulative self self total
> time seconds seconds calls ms/call ms/call name
> 3.37 0.32 0.32 33228 0.01 0.01 headersEnd
> 3.16 0.62 0.30 memPoolGet
> 2.53 0.86 0.24 199695 0.00 0.00 httpHeaderIdByName
> 2.11 1.06 0.20 xprof_stop
> 1.48 1.20 0.14 xprof_start
> 1.37 1.33 0.13 2692928 0.00 0.00 httpHeaderGetEntry
> 1.16 1.44 0.11 351261 0.00 0.00
> statHistBin(_StatHist const*, double)
> 1.11 1.54 0.10 519896 0.00 0.00
> cbdataInternalAlloc
> 1.05 1.65 0.10 memPoolFree
>
> read into this what you will :}.
>
> What I get from it, is that the greatest single improvement we can get
> is only 3%. However, the interesting ration is the self calls / % -
> which is very high for headersEnd.
>
> Rob
>
> --
> GPG key available at: <http://users.bigpond.net.au/robertc/keys.txt>.
>

------------------------------------
 Andres Kroonmaa <andre@online.ee>
 CTO, Microlink Data AS
 Tel: 6501 731, Fax: 6501 725
 Pärnu mnt. 158, Tallinn
 11317 Estonia
Received on Wed Feb 19 2003 - 11:48:32 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:19:16 MST