SUMMARY: CFS Performance
From: Parkin Frank - fparki (fparki_at_acxiom.co.uk)
Date: 09/09/03
- Previous message: Maglinger, Paul: "Escape code for Tru64 Unix 5.1a print jobs"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Date: Tue, 09 Sep 2003 17:51:33 +0100 To: "'tru64-unix-managers@ornl.gov'" <tru64-unix-managers@ornl.gov>
Original Posting:
TruCluster v5.1a / v5.1b
We are currently building and testing ES45 clusters utilising EVA storage.
We have noticed that a single "dd" can overwhelm the memory channel
inter-connect when reading/writing to a disk that is not locally served by
CFS e.g. a drop from 70mb/s to 20mb/s.
Has anyone performed any benchmarking and are willing to share?
We are particularly interested in AdvFS performance and whether more memory
will be required to buffer read/writes (when the IC is saturated) to
maintain disk performance.
Summary:
Thanks go to...
Kristoffel Bruneel,
Rich Glazier
Dan Goetzman.
Dr. Udo Grabowski
Martin Petder
Keith Strange
Apologies for the late summary. The DD test was a poor performance benchmark
as it was direct to disk. Using dt and AdvFS produced some better
statistics.
For disk intensive activities the performance penalty of disk writes over
the cluster interconnect can be 0 - 70% depending on the number of parallel
process. The performance penalty also increases with increased load on the
hosting node. Interestingly, it appeared to be more CPU intensive rather
than memory.
For a mixed process (reads, CPU, writes) the penalty was between 0 - 20%.
Using v5.1b will improve matters as you have the option to dynamically
migrate disks to the node where usage is heavy.
Regards
Frank
**********************************************************************
The information contained in this communication is
confidential, is intended only for the use of the recipient
named above, and may be legally privileged.
If the reader of this message is not the intended
recipient, you are hereby notified that any dissemination,
distribution, or copying of this communication is strictly
prohibited.
If you have received this communication in error,
please re-send this communication to the sender and
delete the original message or any copy of it from your
computer system. Thank You.
- Previous message: Maglinger, Paul: "Escape code for Tru64 Unix 5.1a print jobs"
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Relevant Pages
|
|
