Re: [PROF RFC 2/2] profiling of libcurl by curl-loader, starting stage
Date: Wed, 25 Apr 2007 11:02:08 +0200 (CEST)
On Tue, 24 Apr 2007, Robert Iakobashvili wrote:
> The motivation behind these earlier posted profiling data at the ramp stage
> was to draw attention to the performance issues during connection
> When we attempted to add 100 and more clients each second to load, it has
> completely burned out the CPU. The experiments posted were done with only 20
> clients added each second to enable reliable profiling with oprofile. Both
> gprof and oprofile data show, the CreateConnection () takes a lot of CPU.
I've glanced through the profiling data and I can't see anything significant
that sticks out. Can you?
Have you done any theoretical calculations on what kind of load you should
expect from these operations or do you have anything else to compare with that
would indicate that libcurl is "misbehaving"? And if so, by how much?
Digging in and fully comprehending the profiling data is a big undertaking,
which I'm just not able to do at this point in time.
-- Commercial curl and libcurl Technical Support: http://haxx.se/curl.htmlReceived on 2007-04-25