[Linux-cluster] architecture discussion -- NFS clustering with iSCSI

Rainer Duffner rainer at ultra-secure.de
Mon Jan 28 02:28:44 UTC 2008


Am 28.01.2008 um 02:36 schrieb Terry:

>
>
> Good questions:
> 1) Do you have concurrent writes to the same file from different  
> nodes?
>   1a) No


Well, that's one of the things GFS is good at ;-)


> 2) How many nodes do you have?
>   2a) 3 to start, probably won't go beyond 12
>


OK, that's still in the range GFS can handle (AFAIK).


> I appreciate alternative ideas to NFS.  NFS could possibly introduce
> performance issues (comments here appreciated).


One problem might be that NFS was never supposed to run on  GBit- 
networks.
Thus there is overhead.
But, OTOH, the vendors I mentioned have managed to squeeze a lot of  
performance out of NFS.
It's also a question of optimizing/matching NFS clients and servers.


>   The majority of the
> system is write.  I would say 80%.
>

Do you have a lot of small files?
Small files are usually what degrades GFS-performance.

As I mentioned, if your requirements are tending to be more grid- 
computing related, best ask somebody with a grid-computing background.
Though, they may think in dimensions where 12 nodes is what they have  
at home in a VMware-team.
;-)

Personally, I would be very careful to consider GFS for a system with  
a lot of writes (concurrent to a file or not).
The reason is that it is most times impossible to predict the  
behavior of GFS with a certain application-load-pattern at a given  
cluster-size - there are just too many variables.
This is, of course, also true for NFS - but there's simply much more  
NFS out there "in the field" and you can *always* find someone who  
does nearly (or exactly) the same thing as you want to do and then  
make better educated guesses about how the system will perform.

So, if you want to go with GFS: build your cluster, see if it  
performs, if it doesn't perform: work with your integrator and RedHat  
and see what they can do.
If that doesn't help: scrap it and install an NFS server on some box  
(use Solaris+ZFS - though there are other issues with that, too, of  
course).



cheers,
Rainer
-- 
Rainer Duffner
CISSP, LPI, MCSE
rainer at ultra-secure.de





More information about the Linux-cluster mailing list