kbenson at a-1networks.com
Tue Nov 6 19:13:42 UTC 2007
Chris Johnson wrote:
> On Tue, 6 Nov 2007, Kevan Benson wrote:
>> 1) Stock fuse, or glusterfs patched fuse? See
>> http://ftp.zresearch.com/pub/gluster/glusterfs/fuse/. The Glusterfs
>> team has some changes to some default values in fuse to make it
>> perform better for common glusterfs scenarios, as well as a fix for
>> locking, so you are better off using the glusterfs supplied fuse if
>> you want better performance and or locking.
> Stock because I could get an RPM for it.
If you already have an RPM for 2.7.0, just get the SRPM and modify the
line referencing the package to use from fuse-2.7.0..tar.gz to
fuse-2.7.0-glfs5.tar.gz, and build.
If you aren't familiar with building RPMs, or don't have time, I suggest
removing the RPM for fuse and compiling from source. Not as nice as
keeping everything in an RPM, but it might be worth it.
There's a trade-off here between performance/compatibility and ease of
administration, it's up to you to decide which one's more important.
>> 2) The read-ahead and write-behind translators are there to boost
>> performance for certain scenarios if you know the types of access your
>> mount will be doing much of the time.
> Serial reads and writes mostly. Very little if any random stuff.
I'm not the best person to tell you when or how to use those
translators, but they are there and can probably help.
>> 3) The real speed benefits arise when you are able to span reads
>> across multiple servers, increasing response and transfer rate. This
>> is where the real benefits are (as well as redundancy), which NFS
>> can't really compete with (unless you're using Solaris).
> Striping? I thought that was frowned upon.
Not the stripe translator, but AFR (if and when it supports striped
reading to some degree) and/or unify. If you are disk bound, you could
put up four servers and unify them to achieve a theoretical 4X speedup
in reads and writes (of multiple files) using unify (without AFR
redundancy). You wouldn't necessarily see a speedup in a single read or
write operation, but in most cases that's not what you want to be
>> 4) That's a real close benchmark. Are you sure the medium over which
>> you are transferring the data isn't maxed? IB or TCP/IP? 100Mbit or
>> 1000Mbit (and server grade or workstation grade cards if gigabit).
> The servers are on gigabit. It was a preliminary test. I need
> to do it over known gigabit on both ends. That's next.
I suspect that if your benchmarks are close to the point of 1 ms, you
are seeing a limitation in the disk at either end (are you saving the
file to a local disk, ramdisk or throwing it away?), or a limitation in
the network. I don't usually see a 1 ms difference in concurrent tests
on the same setup, but then I'm not entirely sure I know exactly what
you mean by 1ms/read slower or how you came by that number. Can you
elaborate on your testing?
More information about the Gluster-devel