[patch] giant-less quotas for UFS

[Eric Anderson]

Nicolas KOWALSKI wrote:
> Eric Anderson <anderson at centtech.com> writes:
> 
>> Nicolas KOWALSKI wrote:
>>> Eric Anderson <anderson at centtech.com> writes:
>>>
>>>> Nicolas KOWALSKI wrote:
>>>>> Yes, this is exactly what is happening. To add some precision, some
>>>>> students here use calculation applications
>>>>> that allocate a lot of disk space, ususally more than their allowed
>>>>> home quotas; when by error they launch these apps in their home
>>>>> directories, instead of their workstation dedicated space, it makes
>>>>> the server go to its knees on the NFS client side.
>>>> When you say 'to it's knees' - what do you mean exactly?  How many
>>>> clients do you have, how much memory is on the server, and how many
>>>> nfsd threads are you using?  What kind of load average do you see
>>>> during this (on the server)?
>>> Sorry for the imprecision.
>>> The server is a Dual-Xeon 2.8Ghz, 2GB of RAM, using SCSI3 Ultra320
>>> 76GB disks and controller. It is accessed by NFS from ~100 Unix
>>> (Linux, Solaris) clients, and by Samba from ~15 Windows XP. The
>>> network connection is GB ethernet.
>>> During slowdowns, it's only from a NFS client view that the server
>>> does not respond. For example, a simple 'ls' in my home directory is
>>> almost immediate, but when it slows down, it can take up to 2 minutes.
>>> On the server, the load average goes to 0.5, compared to a default
>>> maximum of 0.15-0.20. The nfsd processus shows them in the state
>>> "biowr" in top, but nothing is really written, because the quotas
>>> system block any further writes to the user exceeding her/his quotas.
>>>
>> In this case (which is what I suspected), try bumping up your nfsd
>> threads to 128.  I set mine very high (I have around 1000 clients),
>> and I can say there aren't really ill-effects besides a bit of memory
>> usage (which you have plenty of).  I suspect increasing the threads
>> will neutralize this problem for you.
> 
> Using 128 nfsd threads, I stressed the server, by running on a NFS
> client a small C program, writting continuously in a file, so that the
> user "biguser" (account stored on /export/home2) exceeds his quota.
> 
> It half-works: during the test, users working on another disk
> (/export/home) did not see any difference, but users working on the
> same disk that "biguser" (/export/home2) where almost halted.
> 
> So, this is better, because before everybody was halted, but there is
> still a problem.
> 
> Any other tips ?

Watch gstat during the testing, and see if the disk that holds the full 
partition is really busy.  I'm betting it's thrashing the disk 
continually checking for free space.  I don't think there's any way to 
avoid that.


Eric




-- 
------------------------------------------------------------------------
Eric Anderson        Sr. Systems Administrator        Centaur Technology
Anything that works is better than anything that doesn't.
------------------------------------------------------------------------