| 1 |
On Fri, Dec 03, 2004 at 10:02:19AM -0900, Leif Sawyer wrote: |
| 2 |
|
| 3 |
> make -j all |
| 4 |
> |
| 5 |
> uptime: 09:41:41 up 10m, 2 users, load avg: 200.90, 65.79, 23.42 |
| 6 |
... |
| 7 |
> Dec 3 09:48:05 VM: killing process apache2 |
| 8 |
|
| 9 |
> OOMkiller 'feels' more aggressive in -rc2, but still doesn't have the |
| 10 |
> instability |
| 11 |
> that gds-267-r16 has |
| 12 |
|
| 13 |
Your workload is pathological. 100+ simultaneous executions of gcc is |
| 14 |
sure to exhaust resources on all but the largest boxes. That the |
| 15 |
silly "oom killer" kills the wrong process and that its behaviour |
| 16 |
varies from kernel to kernel should not be surprising as it has been |
| 17 |
discussed to death on lkml. Don't use it. If you need to keep system |
| 18 |
processes running, disable the oom killer and apply resource limits to |
| 19 |
users. |
| 20 |
|
| 21 |
I question the use of make -j as a stress test. What is the desired |
| 22 |
behaviour of running this command, other than not to crash the OS? If |
| 23 |
you haven't specified any resource allocation policy, the OS's only |
| 24 |
real obligation is not to crash. So what constitutes a "successful" |
| 25 |
test run? |
| 26 |
|
| 27 |
Should the spawned compiler instances fail to allocate memory and |
| 28 |
bomb? Should the OS kill them? Should it kill the pg leader (make) |
| 29 |
instead? Or your shell? Or should some other process's allocations |
| 30 |
fail? Should the OS kill some other process? If so, which one? If |
| 31 |
it's going to kill something, what signal(s) should it force? If it |
| 32 |
sends SIGTERM, what should happen to other processes that attempt to |
| 33 |
sbrk(2) or mmap(2) while it's waiting for the SIGTERM'd process to |
| 34 |
die? Do you put them to sleep? Do you fail their requests? Do you |
| 35 |
kill them too (yay, deadlock!)? What if the process needs to allocate |
| 36 |
memory when dying (more deadlock)? If it sends SIGKILL, what about |
| 37 |
shm segments it may have allocated; wouldn't leaking those just worsen |
| 38 |
the problem? Maybe it should just kill all of userland except init |
| 39 |
and start over. And if it's going to do that, why not just crash? |
| 40 |
|
| 41 |
I don't see that Linux has answered any of these questions, and most |
| 42 |
of them don't need to be asked. Use resource management. If it |
| 43 |
doesn't work, fix it. The tests you are running are a waste of time. |
| 44 |
|
| 45 |
-- |
| 46 |
Keith M Wesolowski |
| 47 |
"Site launched. Many things not yet working." --Hector Urtubia |
| 48 |
|
| 49 |
-- |
| 50 |
gentoo-sparc@g.o mailing list |
Archives