[OpenAFS] openafs performace problems

[Rich Sudlow]

Mattias Pantzare wrote:
> On Tue, Jan 6, 2009 at 20:45, Rich Sudlow <rich@nd.edu> wrote:
>> Mattias Pantzare wrote:
>>> On Tue, Jan 6, 2009 at 17:10, Rich Sudlow <rich@nd.edu> wrote:
>>>> Esther Filderman wrote:
>>>>> To some degree, OpenAFS will always write slower than standard NFS,
>>>>> because AFS is actually making sure it's not writing crap.  NFS will
>>>>> happily write stuff at blazingly fast speeds, not caring whether the
>>>>> data it writes is sane or corrupted.
>>>> The reason NFS appears to be faster is because you're not doing an
>>>> apple - apples comparision - if you were you would have to turn off
>>>> attribute caching on NFS - at that point you'd find that performance
>>>> is essentially equal
>>> Why would you turn off attribute caching? That is a part of NFS.
>> You're correct you generally wouldn't - But if you are truly comparing
>> NFS and OpenAFS you would need to.
>>
>>> Why would attribute caching make the test be an apples - oranges
>>> comparison?
>> Because you have no cache coherancy on NFS to verify that data is propogated
>> out and seen simultaneously on multiple clients
>> (V2 & 3)  whereas with OpenAFS that cache coherancy is there.
> 
> If your existing application is working fine on NFS with attribute
> caching you should _not_ turn it off when you are comparing to AFS.
> The application can clearly work with the attribute cache. (The
> attribute cache is normally not a problem)
> 
> In fact the test would be misleading if you did turn the cache off.

That's correct and then you're comparing apples and oranges.



-- 
Rich Sudlow
University of Notre Dame
Center for Research Computing
128 Information Technology Center
PO Box 539
Notre Dame, IN 46556-0539

(574) 631-7258 office phone
(574) 631-9283 office fax