[Xcb] profiling and performance

[Jamey Sharp]

On Fri, Apr 28, 2006 at 07:00:45PM +0200, Vincent Torri wrote:
> ... my last test with evas shown that xlib was
> a bit faster than xcb. Jamey told me that it was not surprising, as xcb
> has not been intensively profiled. My questions are : can we gain much
> with profiling xcb in that area ? And what should be profiled ?

First you should retest with current XCB: I've committed a number of
performance fixes over the past couple of months. I don't really expect
the performance difference between Xlib and XCB to have an impact on
real applications at this point, though I intend to keep seeking
performance improvements as I have time.

If you profile your test app, and XCB is really the cause of the
performance difference, then I expect you'll find that memcpy and
XCBSendRequest are the most significant users of CPU time. If you find
some other part of XCB is the limiting factor on your application's
performance then I'd like to know.

It's also conceivable that there's a bug in your translation of the
Xlib-based evas code to XCB. If you're issuing more requests than
before, performance would suffer; if you're missing some XCBFlush calls
then response time might be perceptably slower.

There are lots of aspects of performance -- throughput, latency -- and
it would help if you can quantify more precisely how Xlib-based evas is
"a bit faster" than XCB-based evas.

I hope that answers that question. :-)

--Jamey
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 191 bytes
Desc: Digital signature
Url : http://lists.freedesktop.org/archives/xcb/attachments/20060428/a0428e3c/attachment.pgp