[OpenAFS] Speed difference between OpenAFS 1.4.x on Debian and
Mon, 07 Apr 2008 00:06:01 +0200
For starters: I'm replying to the list - maybe somebody would be interested.
Sergio Gelato, dnia 2008-04-06 21:20 napisal:
> * Michał Droździewicz [2008-04-06 10:07:18 +0200]:
>> First I've tried to install OpenAFS with Debian 4.0 (etch) on 3
>> different machines (beginning from old Celeron, through Pentium 4 and =
>> the end on Xeon 3GHz). Speed was pretty much the same when coping file=
>> from local disk to AFS (transfer without the network). It was about 8 =
>> 9 MiB/s and it didn't depend on machine RAM memory nor processor power=
>> Tuning the afs client haven't helped at all (speed was rather dropping=
>> than going up).
> Local disk to AFS isn't a very interesting use case for most people
> since AFS fileservers tend to be dedicated machines. Some people don't
> even install an AFS client on their fileservers. Besides, if there is
> a bottleneck it would be nice to know whether it's on the server or on
> the client side, and for that the tests over a Gigabit network are
> probably best. Then you could even test a Debian server with a CentOS
> client and vice-versa.
Local disk to AFS is interesting for me as a benchmark and as a
restoring client data from back up, because when machine fails you have
to pump it up to the AFS structure somehow, over a network or from a
local disk (in both cases using AFS client). Doing a backup also
requires local reading. Unless there is better way of doing/restoring
Bottleneck is on the server side (as far as I've tested it). When
testing server (both on Debian and on CentOS) besides of the local AFS
client there were one Ubuntu client and two Fedora 7 clients. Every
client had small disk cache and I was testing write using large file (10
GiB) with mc (not very sophisticated, but the same in every try).
Using Debian as operating system, local AFS client was transferring data
@ max speed of 20MiB when alone. When other client connected, transfer
rate was divided equally between two clients. When network client was
transferring data, on 100Mib network it was 7MiB and on 1000Mib network
it was 12-15MiB @ max.
Using Centos local AFS client transfer speed was 38-40MiB, but when I've
connected three clients (one Ubuntu and 2 Fedora 7) on the 100Mib
network they divided bandwidth quite equally, but local AFS client
transfer speed wasn't affected.
Top speed for disk write was 120MiB (two SATA disks in software RAID 0)
in both cases - Debian and CentOS.
>> Lets say that the performance was like 250% better with 1.4.6 on CentO=
>> 5 than with 1.4.2 on Debian 4.0. I've even compiled 1.4.6 and 1.5.34 o=
>> Debian 4.0 but performance was the same.
> I'd move to 1.4.6 on etch in any case.
I like to have software installed from packages, so first I have to
build 1.4.6 for etch and then install it ;)