[Lustre-discuss] MDS performance

[Francois Chassaing]

Hi list, 
I know this question has been asked zillion times, but I'll ask it again, because as most lustre-builder wannabe, I'm worried about MDS being a bottleneck. 
I'm planning to build a 60TB lustre install separated in two pools : 
- one big / slow / cheap 50 TB archive based on Dell MD3000i+MD1000 RAID 5 OST with two cheap OSS in front of it 
- one beefier (=smaller / faster / more expensive) 10 TB based on 2 servers being both OSS+OST each composed of 8 cores, 12 GB RAM, 8 NL SAS spindles RAID 10. 
the big / slow archive is not really a problem but i'm asking myself about the smaller one, given the fact that this storage (if good enough in terms of performance) could double pretty soon by adding more OSS+OST 
I've been planning to have one redundant MDS over 2 Dell R610 sharing a common sas-attached MD3000 RAID 10 MDT. 
I've noticed that this array is not the best on the market for perf. but is rather good at $$... 
I plan on connecting all this on DDR infiniband (along with the main clients -4 of them-). 

So now for my questions about MDS : 
- Should I do better having two lustre installs instead of a single install with two pools ? 
- Should I consider buying a better array for MDT ? 
- Should I be better using even beefier MDS with internal storage and DRBD ? 

and about OSS : 
- should I take "smaller" OSS+OST to improve perf ? 
- should I split my OST storage "inside" the OSSes for the 10GB storage pool ? 

I'll be sooo grateful if someone could answer those, I'll be glad to provide any other details one would need to help me out. 

Thanks. 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20091130/2c9225f0/attachment.htm>