[PVFS2-developers] runaway aio thread

[Pete Wyckoff]

poznick at conwaycorp.net wrote on Wed, 07 Jul 2004 17:50 -0500:
> Just a note, I set up a single-node PVFS2 volume on a machine running
> RHEL3 with update 2, and I don't see this behavior.  When I did a
> pvfs2-cp of a file into the volume, the leaders in CPU usage were
> pvfs2-cp, pvfs2-server, but keventd (I guess that's what handles the aio
> calls into the kernel) was leading the pack (~25%).  It only took 11
> seconds to copy the ~500mb file in, and once that was done the CPU usage
> of everything involved dropped back to zero.
> 
> kernel-smp-2.4.21-15.EL
> glibc-2.3.2-95.20

Thanks for the comparison.  But I just tested again on the latest cvs
here and found the same behavior previously reported:  tight loop in
handle_kernel_aio() in librt.so, part of glibc, when compiled with
-D__PVFS2_TROVE_AIO_THREADED__.  No problem without.

But we have the older glibc-2.3.2-95.6, and a 2.6.6 kernel.  I'll
wheedle our sysadmin to see if we can get some update action on this
machine and try again.

Suggest we table the issue until said time, or until more complaints
crop up, at which time we can follow neillm's suggestion of checking
for precise versions on faulty systems.

		-- Pete