xfs
[Top] [All Lists]

Re: Performance drop in Linux 2.4.19-XFS 1.2 ?

To: Ravi Wijayaratne <ravi_wija@xxxxxxxxx>
Subject: Re: Performance drop in Linux 2.4.19-XFS 1.2 ?
From: Nathan Scott <nathans@xxxxxxx>
Date: Wed, 30 Jul 2003 11:58:54 +1000
Cc: linux-xfs@xxxxxxxxxxx
In-reply-to: <20030730013103.67757.qmail@xxxxxxxxxxxxxxxxxxxxxxx>
References: <20030730013103.67757.qmail@xxxxxxxxxxxxxxxxxxxxxxx>
Sender: linux-xfs-bounce@xxxxxxxxxxx
User-agent: Mutt/1.5.3i
On Tue, Jul 29, 2003 at 06:31:03PM -0700, Ravi Wijayaratne wrote:
> Hi,
> 
> We are seeing a %15 performance drop when we move from XFS 1.1 to 1.2.
> Here are some of our particulars:
> We have been testing the performance of Linux-2.4.19 and XFS 1.2 with 
> NetBench.
> We compared the performance with that of Linux-2.4.18 and XFS 1.1.
> We have been running Samba 3.0.
> 
> 
> The following are the configuration settings:
> 
> RAID config 0 or 5: 4 drives with chunk size of 64
> 
> xfs_info.sh /hd/vol_mnt0/
> meta-data=/hd/vol_mnt0           isize=2048   agcount=74, agsize=1048560 blks
> data     =                       bsize=4096   blocks=76644352, imaxpct=25
>          =                       sunit=16     swidth=16 blks, unwritten=0
> naming   =version 2              bsize=4096  
> log      =internal               bsize=4096   blocks=9360 version=2
>          =                       sunit=16 blks
> realtime =none                   extsz=65536  blocks=0, rtextents=0
> 
> We are running software RAID 0 and 5 (md). The performance numbers were 
> obtained after the
> RAID sync was completed.
> 
> For Linux 2.4.18-XFS-1.1 case we used log version 1.  
> For Linux 2.4.19-XFS-1.2 case we used log version 2.
> 
> At mount time we changed the internal log size from 64k-256k, but that change 
> did
> not show any difference.
> 

Try changing one thing at a time to try figure out what actually
causes the performance changes.  ie. is it the change from 2.4.18
to 2.4.19, or 1.1 to 1.2 XFS code, or is it the use of v1 vs. v2
logs, or is it the use of larger iclogs, or the use of a large log
sunit, or...  just too many variables here.

Posting the results of the benchmarks would be of use too.

thanks.

-- 
Nathan


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