I decided to do a little bit of playing around today and did some performance benchmarks on our eStorage system I’ve been setting up. eStorage is supposed to stand for Enterprise Storage, as opposed to isolated islands of storage. Old storage was DDN 9550’s and the new storage is DDN 9900 Data luns with EMC CX-380 metadata luns. See if you can figure out which is which. (-: Oh, and the drop in performance on the ops2 filesystem is because of a different block size. 256k instead of 128k, need to research why the hit in performance.


According to some documentation forwarded to me the usable throughput of a 4GB fibre connection is 400 MB/s. We’re running ~390 MB/s from the I/O node and ~325 MB/s across our IB connections using GPFS. Not too shabby. Not too shabby at all.

This is the pointy haired boss graph. Simple and shows how we did good spending the money we spent.

Below is a graph I find more useful. It’s a simplified one made from iozone.This is only one file system, and one test out of 4 I normally run. A lot of information in a quick chart. I can gather a ton of data in a hurry with IOZone.

iozone output

iozone output

This graph shows the throughput for different block sizing writing out to different final file sizes. This shows me where I have great throughput for small block sizes that match cache strides, and where  I actually fall out of cache and hit a more stable spindle limit. Fun fun. The empty space on the lower right is because I don’t run small block sizes on larger file sizes. Takes too long.

Bored yet? This is my world. (-; Crazy stuff.