[lwhatley.ctr@navo.hpc.mil: Re: [Pvfs2-developers] Re: [Pvfs2-users] PVFS2 on Infiniband]

[Pete Wyckoff]

kraken at wang-fu.org wrote on Tue, 06 Jun 2006 16:23 -0500:
> Thus spake Pete Wyckoff:
> > lwhatley.ctr at navo.hpc.mil wrote on Mon, 05 Jun 2006 10:53 -0500:
> > > Pete, I think you're on to something!  Adding the sched_yield() call to 
> > > the BMI_ib_testcontext() function has sped things up tremendously.  In 
> > > addition, pvfs2-server now only shows 10% CPU usage (and not for nearly 
> > > as long as previously) in top when doing the pvfs2-mkdir command.
> 
> Of course, in 2.6 kernels, use of sched_yield() means that you may end
> up sleeping a long time, and could be asking for trouble...
> 
> - The behavior of sched_yield() changed a lot.  A task that uses
>   this system call should now expect to sleep for possibly a very
>   long time.  Tasks that do not really desire to give up the
>   processor for a while should probably not make heavy use of this
>   function.  Unfortunately, some GUI programs (like Open Office)
>   do make excessive use of this call and under load their
>   performance is poor.  It seems this new 2.6 behavior is optimal
>   but some user-space applications may need fixing.

Thanks, I vaguely recall that being a bad idea.  I'll keep it out of
the production source.  At some point we can have Lee test
event-driven rather than polling bmi_ib, as it seems we can't trust
the thread scheduler to make sure everything makes progress, at
least on that particular kernel.

		-- Pete