The buildbot now uses ccache, which sped up builds tremendously. Cycle time of build slaves with ccache using is 10-11 minutes; the e8400 and q9300 are almost as fast as the i7 now. (Dunno about the celeron yet.)
A few more flaky tests are now blacklisted, and the cluster has been running all afternoon without a glitch.
I hope to turn on email notifications sometime this week.
On Mon, Aug 29, 2011 at 05:53, Dan Kegel dank@kegel.com wrote:
The buildbot now uses ccache, which sped up builds tremendously. Cycle time of build slaves with ccache using is 10-11 minutes;
Told you ;) From 14 mins to 10-11 mins, nice 25% speed up Although it may make things a bit slower in general for perfectly working patches at first try (extra IO to fill in the cache). However, if/once it's integrated with testbot, it should help even more (especially on slow CPUs)
Frédéric
011/8/29 Frédéric Delanoy frederic.delanoy@gmail.com:
On Mon, Aug 29, 2011 at 05:53, Dan Kegel dank@kegel.com wrote:
The buildbot now uses ccache, which sped up builds tremendously. Cycle time of build slaves with ccache using is 10-11 minutes;
Told you ;) From 14 mins to 10-11 mins, nice 25% speed up Although it may make things a bit slower in general for perfectly working patches at first try (extra IO to fill in the cache). However, if/once it's integrated with testbot, it should help even more (especially on slow CPUs)
Yes, first build is about a minute slower with ccache, but there are so many cache hits on later runs that it's well worth it. And slower CPUs totally love ccache; for the e8400, cycle time went from 17 minutes to 11 minutes. (I'm looking forward to seeing how much it helps the celeron.)
I've made two more changes that bring the cycle time down by 3 minutes, to 7-8 minutes on e8400/e9300/i7 slaves: 1) blacklist the slowest test, wininet/ftp, which took 1 minute all on its own 2) run the headless subset of tests in the background, in their own wineprefix, with DISPLAY unset. These tests don't need DISPLAY. This saves another 2 minutes.
I'd say 7-8 minutes is fast enough for now. Being able to pump out a full build-and-test every 2-3 minutes from my little three-node cluster, with only 8 minutes latency, is pretty cool. - Dan
2011/8/29 Dan Kegel dank@kegel.com:
011/8/29 Frédéric Delanoy frederic.delanoy@gmail.com:
On Mon, Aug 29, 2011 at 05:53, Dan Kegel dank@kegel.com wrote:
The buildbot now uses ccache, which sped up builds tremendously. Cycle time of build slaves with ccache using is 10-11 minutes;
Told you ;) From 14 mins to 10-11 mins, nice 25% speed up Although it may make things a bit slower in general for perfectly working patches at first try (extra IO to fill in the cache). However, if/once it's integrated with testbot, it should help even more (especially on slow CPUs)
Yes, first build is about a minute slower with ccache, but there are so many cache hits on later runs that it's well worth it. And slower CPUs totally love ccache; for the e8400, cycle time went from 17 minutes to 11 minutes. (I'm looking forward to seeing how much it helps the celeron.)
I've made two more changes that bring the cycle time down by 3 minutes, to 7-8 minutes on e8400/e9300/i7 slaves:
- blacklist the slowest test, wininet/ftp, which took 1 minute all on its own
- run the headless subset of tests in the background, in their own
wineprefix, with DISPLAY unset. These tests don't need DISPLAY. This saves another 2 minutes.
I'd say 7-8 minutes is fast enough for now. Being able to pump out a full build-and-test every 2-3 minutes from my little three-node cluster, with only 8 minutes latency, is pretty cool.
- Dan
Also are you doing "./configure -C"? That could save you a few more minutes.
Damjan
On Sun, Aug 28, 2011 at 8:53 PM, Dan Kegel dank@kegel.com wrote:
I hope to turn on email notifications sometime this week.
Buildbot is now sending email on failure. There are two caveats: - it's emailing directly from a cable modem, not using an smtp relay, so some mail systems may think it's spam - buildbot currently explodes while attaching patches to status emails if the patch contains a non-ascii character, http://trac.buildbot.net/ticket/2091 (I'll probably work around this by disabling the attachment.)
Future status updates will go to http://wiki.winehq.org/BuildBot - Dan
On Mon, Aug 29, 2011 at 23:41, Dan Kegel dank@kegel.com wrote:
On Sun, Aug 28, 2011 at 8:53 PM, Dan Kegel dank@kegel.com wrote:
I hope to turn on email notifications sometime this week.
Buildbot is now sending email on failure. There are two caveats:
- it's emailing directly from a cable modem, not using an smtp relay,
so some mail systems may think it's spam
- buildbot currently explodes while attaching patches to status
emails if the patch contains a non-ascii character, http://trac.buildbot.net/ticket/2091 (I'll probably work around this by disabling the attachment.)
Alternatively, avoid attaching patches containing .po or .rc files
2011/8/29 Frédéric Delanoy frederic.delanoy@gmail.com:
- buildbot currently explodes while attaching patches to status
emails if the patch contains a non-ascii character, http://trac.buildbot.net/ticket/2091 (I'll probably work around this by disabling the attachment.)
Alternatively, avoid attaching patches containing .po or .rc files
The patch that caused the problem choked on the first line, From: name-with-accent So it's not just .po files that cause the problem.
I've disabled the attachment code, and am tweaking the message formatting now. Apologies for spamming patch authors with buildbot status results... the current buildbot crashes if you don't send results to the author, or I'd disable the messages. - Dan
-
Damjan Jovanovic -
Dan Kegel -
Frédéric Delanoy