[PVFS-users] pvfs performance analysis

[Martin Margo]

I got a couple of questions  regarding PVFS performance.

my settings:
20 servers, each has a dual 1.5 Ghz Madison (ia64) processors
running suse sles8
kernel 2.4.21_144
running pvfs 1.6.2
disk configuration, each node has dual 73 GB IBM SCSI drives,
	PVFS is setup on the second drive.

we tried using two configurations

configuration 1:
	1 node as mgr,iod,client
	12 nodes as iod and client
	7 nodes as iod

configuration 2:	
	1 node as mgr, iod
	1 node as iod
	12 nodes as clients
	


1. Based on my observation using IOR to benchmark PVFS (2 cases, case1
with 20 iod, case 2 with only 2 iod) i 	saw that the write performance
is roughly half of the read performance. This seems normal but do you
know exactly what is causing this?
2. Based on the same settings as above, my read performance dropped
significantly for I/O greater than 256 MB for configuration 2 and 1 GB 
for
configuration 1. Now this number (256mb, and 1 gb are per node). We
used 12 clients for these tests, so the performance is dropping sharply
after aggregate of 3 GB at configuration 2 and after 12 GB aggregate on 
configuration 1. Does it have
anything to do with congestion, load, etc? Is this common?
3. PVFS read performance dropped again when I did my scaling test (each
client write 1 GB  unique file to pvfs in configuration 2. I varied the
clients from 2 to 12 nodes, and read performance decreased sharply
after 6 clients from 200MB/s into 50 MB/s aggregate.

Thanks in advance.

-Martin W. Margo
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: text/enriched
Size: 1573 bytes
Desc: not available
Url : http://www.beowulf-underground.org/pipermail/pvfs-users/attachments/20040419/20f4246a/attachment.bin