PVFS Tuning
PVFS Development Team
The default settings for PVFS (those provided and in the source code
and added to the config files by pvfs2-genconfig)
provide good performance on most systems and for a wide variety of workloads.
This document describes system level and PVFS specific parameters
that can be tuned to improve
performance on specific hardware, or for specific workloads and usage
scenarios.
In general performance tuning should begin with the available hardware
and system software, to maximize the bandwidth of the network and
transfer rates of the storage hardware. From there, PVFS server
parameters can be tuned to improve performance of the
entire system, especially
if specific usage scenarios are targetted. Finally, file system
extended attributes and hints can be tweaked by different users to
improve individual performance within a system with varying workloads.
Some (especially system level) parameters can be tuned to provide better
performance without sacrificing another property of the system.
Tuning some parameters though, may have a direct effect on the
performance of other usage scenarios, or some other property of the
system (such as durability).
Our discussion of performance tuning will include the tradeoffs
that must be made during the tuning process, but the final decisions are best
made by the administrators to determine the optimal
setup that meets the needs of their users.
For users that have one use case, and a generic cluster, what's the best
partition of compute/IO nodes? Is this section needed?
How many IO servers? 1How many MD servers?
Should IO and MD servers be shared?
The easiest way to see an improvement in performance is to set the
TroveSyncMeta and TroveSyncData attributes to ``no''
in the <StorageHints> section. If those attributes are set to
``no'' then Trove will read and write data from a cache and not the
underlying file. Performance will increase greatly, but if the server
dies at some point, you could lose data. At this point in PVFS2
development, server crashes are rare outside of hardware failures.
PVFS2 developers should probably leave these settings to ``yes''. If
PVFS2 hosts the only copy of your data, leave these settings to ``yes''.
Otherwise, give ``no'' a shot.
Sync or not, metadata, data
coalescing
distributed metadata
- Unexpected message size
- Flow Parameters
- Number of Datafiles
- Stripe Size
- Distribution
PVFS Tuning
This document was generated using the
LaTeX2HTML translator Version 2002 (1.62)
Copyright © 1993, 1994, 1995, 1996,
Nikos Drakos,
Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999,
Ross Moore,
Mathematics Department, Macquarie University, Sydney.
The command line arguments were:
latex2html -split 0 -show_section_numbers -nonavigation -init_file /tmp/pvfs-2.7.1/doc/latex2html-init pvfs2-tuning.tex
The translation was initiated by Samuel Lang (ANL) on 2008-04-14
Footnotes
- ... servers?1
- The FAQ already answers this to some
degree
Samuel Lang (ANL)
2008-04-14
|