problem with nfs ls latency during high IO

Judith Flo Gaya jflo at imppc.org
Tue Mar 15 16:49:17 UTC 2011



On 03/15/2011 03:52 PM, Michael Cronenworth wrote:
>
> I wouldn't know the exact answer to that question, but the kernel
> changes very frequently. RHEL5 is 17 versions behind F14. RHEL6 is
> already 3 behind F14. If you want to be sure that the NFS code is all
> identical then you should ask on the kernel list or an NFS list.
I did, had no answer by the moment.
I asked here because I thought that maybe there was some kind of issue 
fedora-related or maybe someone had the same problem here..
>
>
> There might be different kernel defaults for RHEL6. I have not studied
> the two distributions that closely.
The mount parameters look the same according to /proc/mounts, at least 
regarding nfs. The rest of the kernel I don't know.
>
> However, 40 seconds does not seem ideal to me either.
comparing to the 2,3 or even 6 minutes that I usually get from fedora 
while doing one single transfer...
The 40 seconds that it takes from RHEL6 during the start up of the 
second transfer looks ok to me...
> Have you
> considered it may be an issue with your IBRIX?
Yes, in fact I was sure about this and to prove it (considering that hp 
only supports RH), I installed a machine with RHEL6 and did the same 
test with the results that I commented.
> I'm not familiar with
> those devices to make a suggestion though. I ran a simple test with 4.5
> gig file and continuously listed the destination directory from the
> client during the transfer.
So did I, but my file was 36Gb, while doing several ls -l in the red 
hat, I could only do 2 or 3 petitions within the fedora because of the 
hangs ;(
Both tests were done at the same time and there is exactly the same 
network hardware in the middle of the two (1 switch and the network 
cards are 1gbit for both machines) and the copy time was also the same 
(6 minutes and some seconds).
> It took at most 5 seconds to see the listing
> out of 6 listing attempts. I was getting about 100mb/sec throughput
> (1gbit network). Fedora 14 server and Fedora 14 client.
In my case, the servers are RHEL5, one of the clients is fedora 14 and 
the other is RHEL6, all of them 64bit.
I also made some test using a fedora 10 server and a fedora 14 client 
and the times were also big, that's why I asked here.

Again, thanks for your help, it is really appreciated.


More information about the users mailing list