[Swift-devel] GPFS issue of SWIFT on BGP

[Zhao Zhang]

Hi, Ben

What if there are 640 ssh-providers sending result files at the same 
time? Do you know any successful test case
with hundreds of ssh-providers working together with one submit host?

zhao

Ben Clifford wrote:
> On Tue, 10 Feb 2009, Zhao Zhang wrote:
>
>   
>>>> Here we got an old problem, all IO nodes are trying
>>>>         
>
>   
>> data transfer from  I/O nodes to submit-side client is  writing to GPFS
>> through ssh-provider.
>>     
>
> By 'old problem' I assumed you meant the GPFS locking problems previously 
> experienced, where GPFS locks for particular filesystem objects would 
> need to be expensively moved between nodes.
>
> However, that should not be a problem here - from the GPFS perspective, 
> the submit (login) node is the only node that is interacting with GPFS, 
> and the only node that needs a lock on that directory.
>
> If you're experiencing slowness, then I would be inclined to investigate 
> elsewhere. It may be that the ssh provider is not fast (ssh is not 
> renowned for being a fast protocol; Mihael might have some commentary 
> either way based on his experiences with the cog ssh provider); it may be 
> that something else is causing a bottleneck.
>
> Do you have any detailed timing information? (from my perspective, the 
> wrapper logs for every job, and the submit side log, would be interesting 
> to look at - send those to me)
>
>