Dear list, We''re currently experimenting with Lustre and small files (<=1MB) and some questions came to mind: - will writing files that are smaller than the stripe_size and that are written to a fully striped directory create objects on more than 1 OST? - what would be the main options to tune in order get the best out of Lustre for small files? (If the answer to the 1st question is ''yes'', then ''avoid striping'' would be a first candidate, I guess ;) TIA, Arne -- Dr. Arne Wiebalck CERN -------------- next part -------------- A non-text attachment was scrubbed... Name: smime.p7s Type: application/x-pkcs7-signature Size: 6380 bytes Desc: S/MIME Cryptographic Signature Url : http://lists.lustre.org/pipermail/lustre-discuss/attachments/20091030/d49d943d/attachment.bin
[ ... ]> - will writing files that are smaller than the stripe_size and > that are written to a fully striped directory create objects > on more than 1 OST?Let''s say that in some previous discussion one of the Sun/CFS people mentioned storing collections of small files into ''ext2'' image files on Lustre and then loopback mounting those on clients. While I have mentioned using DB4 or similar indexed databases.> - what would be the main options to tune in order get the best > out of Lustre for small files? [ ... ]That depends a lot on the specific access patterns, and what you mean by "small". I guess, also from reading some previous comments on the topic, that the first thing to consider is appropriate storage backends (e.g. SSD, and Sun/CFS otherwise recommend RAID10) for the MDTs, and maybe even for the OSTs. Lots of small files, lots of IOPs (usually).
On 2009-10-30, at 05:21, Arne Wiebalck wrote:> We''re currently experimenting with Lustre and small files (<=1MB) > and some questions came to mind: > > - will writing files that are smaller than the stripe_size and that > are written to a fully striped directory create objects on more > than 1 OST?The striping is set at file creation time, before any information is known about how much data will go into the file.> - what would be the main options to tune in order get the best out > of Lustre for small files? (If the answer to the 1st question is > ''yes'', then ''avoid striping'' would be a first candidate, I guess ;)Right - wide striping should only be used when it is known that the file size will be very large and/or accessed by many nodes at one time. Secondly, using RAID-1+0 OSTs instead of RAID-5/6 will definitely also improve performance, since there is RAID parity overhead for writing small chunks of data to disk. Cheers, Andreas -- Andreas Dilger Sr. Staff Engineer, Lustre Group Sun Microsystems of Canada, Inc.