[Pvfs2-developers] Coalesce patch

Sam Lang slang at mcs.anl.gov
Wed Jul 12 17:21:38 EDT 2006


On Jul 12, 2006, at 3:44 PM, Julian Martin Kunkel wrote:

> Hi,
> Thanks, hope the patch works alright...

Can you change the immediate completion flag and send another?

>>  I'm not convinced this is the best value for the default case.
>> If the server can handle operations fast enough that the queue
>> remains at 1, coalescing doesn't seem necessary, and it will give
>> widely varied timings for some of the operations of some clients.
> I think in general this is a good value, but change it if you like.
>
> If the disk is fast enough that the queue remains at 1 the loss is  
> small
> (depending on the processing time), however you are right it slows  
> the first
> one down.
>
> Argument for setting it to 1:
> We want to coalesce as much ops as possible, assume each sync costs  
> 5 ms (disk
> access time) while a non-blocking modification might take 0.1ms.
> If somebody has 2 ops in the queue and the low watermark is 2 the  
> first
> finishes on 5.1 ms, while the second finishes on 10.2 ms.

No if there are 2 ops in the queue and the low watermark is set to 2  
than they get coalesced.  The case you're probably looking for is if  
the low watermark is set to two, and the queue never gets higher than  
1, but the second operation is added to the queue while the first is  
being serviced.  Its true that the 2 consecutive syncs will take ~ 10  
ms total instead of something like ~ 7 ms, but at least the first  
operation isn't taking 2 ms longer.  Remember that this specific  
scenario means that the client is only sending requests at a rate of  
~ 5 ms, so individual syncs don't really cost anything.

> In case the watermark is 1 the first and the second one finishs at  
> 5.2 ms.
> Even if you assume that the op might be blocking I would say for  
> the majority
> of people setting it to 1 reduce the overall op time.
>
> julian
>
> _______________________________________________
> Pvfs2-developers mailing list
> Pvfs2-developers at beowulf-underground.org
> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
>



More information about the Pvfs2-developers mailing list