On 5/22/19 6:10 AM, Gerald Henriksen wrote:
Paper on arXiv that may be of interest to some as it may be where HPC
is heading even for private clusters:
In case it's of interest NERSC has a page on how Shifter does containers
and how it packs filesystems to improve performance here:
https://docs.nersc.gov/programming/shifter/overview/
That links to the Cray User Group paper and presentation from 2015, but
the page has a more recent graph illustrating how much less time it
takes to run the Pynamic benchmark at 4,800 ranks with Shifter versus
normally using different filesystem options.
The reason is (as the page says):
# Shifter mounts the flattened image via a loop mount. This approach
# has the advantage of moving metadata operations (like file lookup)
# to the compute node, rather than relying on the central metadata
# servers of the parallel filesystem. Based on benchmarking using
# the pynamic benchmark, this approach greatly improves the
# performance of applications and languages like Python that rely
# heavily on loading shared libraries Fig. 2. These tests indicate
# that Shifter essentially matches the performance of a single
# docker instance running on a workstation despite the fact that
# shifter images are stored on a parallel filesystem.
Full disclosure: I'm working at NERSC now (though this all predates me!)
All the best,
Chris
--
Chris Samuel : http://www.csamuel.org/ : Berkeley, CA, USA
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org sponsored by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit
https://beowulf.org/cgi-bin/mailman/listinfo/beowulf
