On 03/13/2015 01:58 PM, Kevin Fenzi wrote:
On Fri, 13 Mar 2015 07:46:04 +0100
Ralf Corsepius <rc040203@xxxxxxxxxx> wrote:
On 03/12/2015 05:06 PM, Ralf Corsepius wrote:
Well, all I can say, building rawhide packages (esp. C++) seems to
have evolved into a lottery and rendered working one Fedora
packages into an adventure.
E.g. I am facing packages which one day build, but fail to build the
other day. Today, I have been facing a package which builds in local
rawhide mocks, but fails in official koji.
Let me provide some facts, as they happened to me, yesterday
(2015-03-12): Trying to build k3d[1] for rawhide
- ca. 10:00 UTC: local x86_64-mockbuild succeeds
- 10:15 UTC: Building in koji
koji fails on x86_64: c++: internal compiler error: Killed
(program cc1plus)
https://koji.fedoraproject.org/koji/taskinfo?taskID=9208659
builder: buildhw-06
Mar 12 10:22:17 buildhw-06.phx2.fedoraproject.org kernel: Out of memory: Kill process 10227 (cc1plus) score 29 or sacrifice child
Mar 12 10:22:17 buildhw-06.phx2.fedoraproject.org kernel: Killed process 10227 (cc1plus) total-vm:635536kB, anon-rss:578384kB, file-rss:17300kB
Mar 12 10:22:17 buildhw-06.phx2.fedoraproject.org kernel: cc1plus: page
allocation failure: order:0, mode:0x280da
- 10:36 UTC: Presuming a temporary hickup, retrying koji
koji fails on x86_64: c++: internal compiler error: Killed
(program cc1plus)
https://koji.fedoraproject.org/koji/taskinfo?taskID=9208790
builder: buildhw-06 again
Mar 12 10:43:46 buildhw-06.phx2.fedoraproject.org kernel: Out of memory: Kill process 13650 (cc1plus) score 29 or sacrifice child
Mar 12 10:43:46 buildhw-06.phx2.fedoraproject.org kernel: Killed process 13650 (cc1plus) total-vm:647752kB, anon-rss:594536kB, file-rss:12628kB
Mar 12 10:43:57 buildhw-06.phx2.fedoraproject.org kernel: as invoked
oom-killer: gfp_mask=0x201da, order=0, oom_score_adj=0
...snip...
?!?
My conclusion is that gcc5 is at least sometimes using a lot more
memory than older versions.
I haven't noticed a significant increase in memory usage in my local
mock builds, but I haven't tried to investigate.
Perhaps reducing smp_mflags could help?
Dunno ;)
buildhw boxes have 20GB of memory and 16CPUs... so perhaps 16 threads
is causing it to hit OOM much more often than in the past?
>
buildvm boxes have 10GB of memory, but only 4 (faster) cpus.
arm boxes have 4GB of memory and 4 cpus.
Does this explain why my yesterday's scratch-builds didn't fail but
"real" builds failed?
Does this explain why "real" builds failed yesterday but succeed today?
Did you change something about this configuration or has something
related to g++ changed within the last 24hrs?
... too many unknowns for my taste.
Ralf
--
devel mailing list
devel@xxxxxxxxxxxxxxxxxxxxxxx
https://admin.fedoraproject.org/mailman/listinfo/devel
Fedora Code of Conduct: http://fedoraproject.org/code-of-conduct
