Thanks Simon and Jason,

thank you for the clues!

we are running openafs 1.4.11 on centos 5.4 using the official openafs rpm 
packages.

I tried running the afsd with -dynroot -fakestat -stat 200000 and still have 
the issue. if I route the traffic (about 500 requests) to that server, it gets 
locked.

how about the -daemons parameter? could that be relevant?

I guess the xstat_cm_test is not part of the RPM packages.

thank you, Nick.

> ------------ Původní zpráva ------------
> Od: Simon Wilkinson <[email protected]>
> Předmět: Re: [OpenAFS] cache manager locked under heavy load?
> Datum: 06.2.2010 22:37:37
> ----------------------------------------
>
> On 6 Feb 2010, at 21:54, Jason Edgecombe wrote:
>
> > I'm not familiar with the error, but the fact that "vcache" is in the 
> > message
> suggests that you vnode/stat cache is not big enough.
>
> The lock that message refers to is in GetVCache. This suggests that you're
> thrashing the vcache - as Jason suggests I'd take a careful look at your cache
> statistics.
>
> I do have some longer term plans for ways of decreasing the contention here.
> Apache really likes the stat() command, and if we could find ways of reducing
> the locking necessary to return results to stat(), web serving from AFS could 
> be
> substantially faster.
>
> On 6 Feb 2010, at 21:43, Russ Allbery wrote:
>
> > We used to see this and then it went away with the current client cache
> > manager.  What version of OpenAFS are you using on your clients?
>
> I suspect that the work that Derrick did on dynamic vcaches will help with 
> this
> particular case, as it should move cache flushes away from the kernel thread
> that's actually doing the access.
>
> Cheers,
>
> Simon._______________________________________________
> OpenAFS-info mailing list
> [email protected]
> https://lists.openafs.org/mailman/listinfo/openafs-info
>
>
>
_______________________________________________
OpenAFS-info mailing list
[email protected]
https://lists.openafs.org/mailman/listinfo/openafs-info

Reply via email to