Hi Hamza,
Running extra I/O applications on the same node is going to result in
decreased apparent performance for each process: you're doing N times as
much with with the same client and network connection. The CPU
scheduler gives each one a fair share of time, so they all end up
getting the processor 1/Nth of the time and so take roughly N times as long.
For the multi-node cases, you did get a performance increase overall in
the "Group 2" case. Three processes wrote data in about the time it
would have taken two to write if they had done so on one processor.
Likewise for reading in "Group 6" case.
Are you using nodes as both servers and clients?
Thanks,
Rob
Hamza KAYA wrote:
When I tried on different machines I got better results but again
worse than one execution. Results are given below. So file access
performances decrease linearly with the number of accesses. Is there a
configuration that will make pvfs2 scale better for concurrent
accesses? Why pvfs2 behave like this?
-- GROUP - 1---
time -p ./write /pvfs2/t1 100 => 26.44 [executed on master]
time -p ./write /pvfs2/t2 100 => 26.72 [executed on master]
time -p ./write /pvfs2/t3 100 => 26.57 [executed on master]
--- GROUP - 2---
time -p ./write /pvfs2/t4 100 => 17.69 [executed on node1]
time -p ./write /pvfs2/t5 100 => 17.55 [executed on node2]
time -p ./write /pvfs2/t6 100 => 16.91 [executed on node3]
--- GROUP - 3---
time -p ./write /pvfs2/t7 100 => 9.42 [executed on master]
--- GROUP - 4---
time -p ./write /pvfs2/t8 100 => 9.12 [executed on node1]
--- GROUP - 5---
time -p ./read /pvfs2/t4 ./t4 => 26.17 [executed on master]
time -p ./read /pvfs2/t5 ./t5 => 27.00 [executed on master]
time -p ./read /pvfs2/t6 ./t6 => 26.92 [executed on master]
--- GROUP - 6---
time -p ./read /pvfs2/t4 ./t4 => 14.64 [executed on node1]
time -p ./read /pvfs2/t5 ./t5 => 16.62 [executed on node2]
time -p ./read /pvfs2/t6 ./t6 => 15.55 [executed on node3]
--- GROUP - 7---
time -p ./read /pvfs2/t4 ./t4 => 9.89 [executed on master]
--- GROUP - 8---
time -p ./read /pvfs2/t4 ./t4 => 10.02 [executed on node1]
On 11/22/05, Rob Ross <[EMAIL PROTECTED]> wrote:
The numbers make absolute sense for four executions on the same machine.
Rob
Hamza KAYA wrote:
Yes. I'll try them on different machines too. I'll inform you as soon
as possible.
Thanks very much.
--
Hamza
On 11/21/05, Rob Ross <[EMAIL PROTECTED]> wrote:
Hi,
Are you running all those processes on the same machine?
Rob
Hamza KAYA wrote:
Thanks very much. Another problem I observed is about simultaneous
accesses. Normally a copying a 100MB file to pvfs takes approximately
10secs. in my system.
time -p ./copy testfile /pvfs2/testfile -> 9.76sec.
However copying 4 files simultaneously gives the following results:
time -p ./copy testfile1 /pvfs2/testfile1 -> approx. 36sec.
time -p ./copy testfile2 /pvfs2/testfile2 -> approx. 36sec.
time -p ./copy testfile3 /pvfs2/testfile3 -> approx. 36sec.
time -p ./copy testfile4 /pvfs2/testfile4 -> approx. 36sec.
[here 'copy' is a programme which uses the system calls. However same
result occurs while using coreutils 'cp' and a simple program which
makes consecutive fread and fwrite calls.]
All of the files used are 100MB. And most of the operations
overlapped. Another point is the CPU consumption of pvfs2-client-co.
When multiple accesses to one file occurs, it consumes approx. %40 of
the CPU.
e.g.
cp /pvfs2/test test1
cp /pvfs2/test test2
cp /pvfs2/test test3
cp /pvfs2/test test4
What may be the problem? Or is this situation is a problem?
Thanks,
--
Hamza
On 11/18/05, Robert Latham <[EMAIL PROTECTED]> wrote:
On Thu, Nov 17, 2005 at 05:08:45PM +0000, Number Cruncher wrote:
I noticed Rob's post at
http://lists.gnu.org/archive/html/bug-coreutils/2005-11/msg00068.html
which discusses a patch to cp. Has this been accepted? Where can I get a
copy (excuse the pun!)
The attached patch, based largely on earlier efforts by Neill Miller,
to coreutils CVS will make copy behave better with pvfs2.
Unfortunately, this patch also modifies lib/Makefile.am, so you'll
need fairly recent versions of autotools/automake/autowhatever. It
should apply ok against coreutils-5.92
I don't know if this will make it into coreutils-6.0, but i'll keep
bugging the maintainers...
==rob
--
Rob Latham
Mathematics and Computer Science Division A215 0178 EA2D B059 8CDF
Argonne National Labs, IL USA B29D F333 664A 4280 315B
_______________________________________________
PVFS2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
_______________________________________________
PVFS2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
_______________________________________________
PVFS2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
