MailChimp Developer

Sign up for our newsletter and get the latest HPC news and analysis.
Send me information from insideHPC:

Dell and Terascala Metadata Performance

dell_blue_rgbThe nice folks at Dell were kind enough to point us to their second installment of the Dell & Terascala technical performance review.  In part two of the series, Scott Collier outlines their recent tests of the Lustre metadata system as a function of creating files and directories in the global file space.

Metadata is basically data about data. We are trying to stress the array and find out how quickly we can create / delete / update and get the status of both files and directories. The test we will be using is “mdtest”. mdtest is an MPI-coordinated benchmark test that performs create / stat / delete operations on files and directories and then reports performance. The version of mdtest that we are using was patched by Li Ou, another HPC engineer on our team. His version includes utime results.

The tests were run on the same configuration as we detailed from their first post.  They report two sets of metadata results.  The first set includes the N-to-N metadata performance of directory creation, directory stat operations, directory utime operations, directory removal, file creation time, file stat operations, file utime operations, and file removal operations.  The tests are scaled across an increasing number of nodes.  Performance seems to peak somewhere between 12 and 24 nodes.

This test shows that directory creates are much faster than file creates. The reason for this is because a file create involves operations on both the MDT and the OST. Directory creates only require operations on the MDT.

The second test stressed the N-to-1 versus N-to-N performance of the MDT.  According to Scott, the N-to-N performance is quite a bit higher than the N-to-1 due to the locks and serialization required to perform N-to-1 operations.

Scott’s post also goes into detail regarding the Dell/Terascala GUI Management console.  Back in the old days, all we had to administer our dusty Lustre file systems was home grown shell scripts.  Oh the joy of grepping proc entries for file system health.

If you’re interested in Lustre performance and configuration, check out Scott’s latest post.  Its a goodie.


  1. Shell scripts ? You were lucky, we never had shell scripts when we were lads. We had to type out the commands every minute on the minute and god help us if we made a mistake…

    (with many apologies to Monty Python)

Resource Links: