xfs
[Top] [All Lists]

Re: fresh xfs vs used xfs

To: David Chinner <dgc@xxxxxxx>
Subject: Re: fresh xfs vs used xfs
From: Deanan <delusion@xxxxxxxxxxxx>
Date: Thu, 15 Sep 2005 09:14:18 -0700
Cc: linux-xfs@xxxxxxxxxxx, sandeen@xxxxxxx
In-reply-to: <20050912224805.GA685356@xxxxxxxxxxxxxxxxx>
References: <Pine.LNX.4.63.0509121822550.3870@xxxxxxxxxxxxxxx> <4325B3C1.20506@xxxxxxxxxxxx> <20050912224805.GA685356@xxxxxxxxxxxxxxxxx>
Sender: linux-xfs-bounce@xxxxxxxxxxx
User-agent: Mozilla Thunderbird 1.0 (Windows/20041206)
Hi Dave & Eric,

Basically I'm trying to write 16MB files at about 400MB/s. The bandwidth
to disk is about 550MB/s so I have a little overhead. On the first pass
writing the files I can make it all the way through the 5.5TB with no problems.
After deleteing the files and starting the second pass, I start seeing delays
in the writes that add up and eventually I run out of buffers.

Unfortunately, I don't have any more info other that the writes start to take
longer.

The nr_requests is already at 128. Should I try something higher?

I'm trying Eric's suggestions now also.

Cheers,

Deanan


On Mon, Sep 12, 2005 at 09:58:41AM -0700, Deanan wrote:
  
Hi,

I'm seeing some strange things happen that i can't really explain.
I'm using the XFS with SLES 9 and 2.6.5-7.97 on x86_64.

I'm filling up 6TB in one shot with 16mb files. When I delete the
directory that contains all the files and then go back the to refill
the directory again, I get performance problems.
    

Can you be more precise about what your performance problem is and
how you are reproducing it?

BTW, if you are using the cfq elevator, what happens if you
"echo 128 > /sys/block/sdXX/queue/nr_requests" for each block device
that makes up your filesystem?

Cheers,

Dave.
  

<Prev in Thread] Current Thread [Next in Thread>