Kyle,

Thanks for the prompt reply.  I had been testing with 10-20GB files.  So today 
I used 100GB files.  Performance was the same for both PVFS and NFS when using 
the 20GB files.   PVFS reads were slower than writes.  NFS reads were faster 
than NFS writes and way faster than PVFS reads.  However, when I tried using 
multiple NFS clients reading the same filesystem (but different files), 
performance of NFS dropped 94%.   PVFS performance only dropped 25%.

I'm already using the alt-io method which appears to be the default setup.

Thanks for your help!
-Roger

-----------------------------------------------------------
Roger V. Moye
Systems Analyst III
XSEDE Campus Champion
University of Texas - MD Anderson Cancer Center
Division of Quantitative Sciences
Pickens Academic Tower - FCT4.6109
Houston, Texas
(713) 792-2134
-----------------------------------------------------------

From: Kyle Schochenmaier [mailto:[email protected]]
Sent: Monday, July 01, 2013 4:02 PM
To: Moye,Roger V
Cc: [email protected]
Subject: Re: [Pvfs2-users] Are writes supposed to be faster than reads?

Hi Roger -

Depending on your setup I would expect that writes would be faster than reads 
simply because you'll have a lot of server-side caching of the data on writes 
and those might be sent out to disk before your reading tests.  Are you doing 
any tests that you're positive you're exceeding the caches on the servers for 
writes with?  (writing out 100GB+ files for example) ?

There are some pvfs2 tunables that will force disk flushes on every write in 
the config file, which would help to see if this really is just caching, Im not 
sure what the flag is off the top of my head though.  I think it is setting 
troveio to 'directio' :

  *   directio. This uses a direct I/O implementation to perform I/O operations 
to datafiles. This method may give significant performance improvement if PVFS 
servers are running over shared storage, especially for large I/O accesses. For 
local storage, including RAID setups, the alt-aio method is recommended.

Also keep in mind that NFS can do client side read caching so reads are almost 
always going to be faster than writes.

Cheers,
~Kyle


Kyle Schochenmaier

On Mon, Jul 1, 2013 at 3:49 PM, Moye,Roger V 
<[email protected]<mailto:[email protected]>> wrote:
We are currently testing a new PVFS2 setup at our facility.  The servers are 
running on RHEL 6 using XFS as the underlying filesystem.  We are using four 
PVFS servers.  Performance is beating NFS, which is what we wanted and expected.

While running some simple I/O tests today I discovered that writes are about 
30% faster than reads.  I expected reads to be faster.   Further, when I tried 
the same I/O tests to the local disk of the client, and to an NFS server 
running on the same server hardware, I found that reads were indeed faster than 
writes by 3:1.

Is it expected that reads would be slower on PVFS?   Is there something I need 
to tune?

Thanks in advance!
-Roger Moye

-----------------------------------------------------------
Roger V. Moye
Systems Analyst III
XSEDE Campus Champion
University of Texas - MD Anderson Cancer Center
Division of Quantitative Sciences
Pickens Academic Tower - FCT4.6109
Houston, Texas
(713) 792-2134<tel:%28713%29%20792-2134>
-----------------------------------------------------------


_______________________________________________
Pvfs2-users mailing list
[email protected]<mailto:[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

Reply via email to