slow OOM killing with 2.6.9?

slow OOM killing with 2.6.9?

[Chris Friesen]

I've got a ppc box with 2GB of ram, running 2.6.9.

If I run a few instances of memory chewing programs, eventually the OOM-killer 
kicks in.  At that point, the machine locks up for about 10 seconds while 
deciding what to kill.

Is this expected behaviour?  Is there any way to speed this up?

Thanks,

Chris

Re: slow OOM killing with 2.6.9?

[Chris Ross]

Chris Friesen escreveu:
> I've got a ppc box with 2GB of ram, running 2.6.9.
> 
> If I run a few instances of memory chewing programs, eventually the 
> OOM-killer kicks in.  At that point, the machine locks up for about 10 
> seconds while deciding what to kill.

OOM killing is known to be broken in 2.6.9, specifically it kills things 
even when the machine isn't out of memeory and/or kills the "wrong" 
things when it is. See threads assim for more details.

The OOM Killer is working properly again in 2.6.10-rc2-mm4. Could you 
try that kernel and report whether it fixed your problems too?

Regards,
Chris R.

Re: slow OOM killing with 2.6.9?

[Chris Friesen]

Chris Ross wrote:
> 
> Chris Friesen escreveu:
> 
>> I've got a ppc box with 2GB of ram, running 2.6.9.
>>
>> If I run a few instances of memory chewing programs, eventually the 
>> OOM-killer kicks in.  At that point, the machine locks up for about 10 
>> seconds while deciding what to kill.
> 
> 
> OOM killing is known to be broken in 2.6.9, specifically it kills things 
> even when the machine isn't out of memeory and/or kills the "wrong" 
> things when it is. See threads assim for more details.
> 
> The OOM Killer is working properly again in 2.6.10-rc2-mm4. Could you 
> try that kernel and report whether it fixed your problems too?

Hmm...downloaded 2.6.9, patched to 2.6.10-rc2, patched to 2.6.10-rc2-mm4.  Tried 
building and got the following error:

[cfriesen@hsdbsk204-83-218-112 linux-2.6.10-rc2-mm4]$ make
   CHK     include/linux/version.h
make[1]: `arch/ppc/kernel/asm-offsets.s' is up to date.
   CHK     include/linux/compile.h
   CHK     usr/initramfs_list
   GEN     .version
   CHK     include/linux/compile.h
   UPD     include/linux/compile.h
   CC      init/version.o
   LD      init/built-in.o
   LD      .tmp_vmlinux1
arch/ppc/mm/built-in.o(.init.text+0x5f4): In function `paging_init':
: undefined reference to `pgd_offset_is_obsolete'
make: *** [.tmp_vmlinux1] Error 1

Any ideas?

Chris

Re: slow OOM killing with 2.6.9?

[Chris Friesen]

Chris Ross wrote:

> The OOM Killer is working properly again in 2.6.10-rc2-mm4. Could you 
> try that kernel and report whether it fixed your problems too?

I had to turn off HIGHMEM support (see other response) but I mangaged to 
reproduce it.

My testcase forks a single child, then both parent and child proceed to allocate 
and write to 3/5 of total system memory.  Swap is disabled.  There is barely 
anything else running, so only one hog should have to be killed.

With 2.6.10-rc2-mm4, I got the following results:


running as root:
trial 1:
both memory hogs as well as the xterm in which they were started were killed. 
In the process, the system was frozen for 14 seconds.

trial 2: both memory hogs killed, system frozen for 9 seconds

trial 3: both memory hogs as well as the xterm in which they were started were 
killed., system was frozen for 15 seconds.

running as regular user:
trial 1:
one hog killed, system hung for 8 seconds

trial 2:
both hogs killed, timing window killed, system hung for at least 15 seconds


Doesn't look like the oom killer is entirely fixed.

Chris