[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: xfs vs jfs performance

On Tue, 11 Feb 2003 22:57, Thedore Knab wrote:
> I am not talking about huge delays but rather occasional 2-5 second delays.

Have you run "iostat -x 10" or similar to see what those delays are from?

> I calculate that Courier IMAP is moving about 200-500 files every minute
> during the delays.
> Additionally, mail is coming in at the rate of 100-300 messages per minute.

I've got ~200,000 users per server on ReiserFS and it's performing as well as 
can be expected.  Generally performance is good enough to make everyone 
happy.  Having >300 POP connections and >40 IMAP connections to a server at 
one time is not uncommon.

> Since ext3 is built on top of ext2, it adds a lot of overhead.

I disagree with that statement.  Ext3 has been shown to increase performance 
over Ext2 in some situations, data=journal is the best example.

Ext3 is not built on top of Ext2, it is a new file system that shares some of 
the same data structures as Ext2 and has backward compatability.

> The kernel quotas add more overhead.

Especially if they are journalled with the related meta-data changes (as is 
required for correct functionality).

Quotas in Ext2 have traditionally been a hack and AFAIK that has not changed.  
The fact that quotacheck is separate from fsck is a sign of this.  XFS wins 
in this regard, and may give you superior performance for this reason.

> Since the system is not being taxed in any other noticeable way
> according to sar, I feel that the file-system must be the bottleneck.
> More specifically, it has to be ext3 or the quotas with ext3.

Traditional sar is not much good.  I recommend compiling your kernel with 
CONFIG_BLK_STATS and using iostat.

Also does your RAID-5 have a battery backed write-back cache?  This makes a 
huge difference to performance.  Onother thing you could do is install a 
non-volatile RAM device such as a UMEM card for the journals and use 
data=journal.  Then the only synchronous writes would be to a 240MB/s device 
with 0 latency.  I've been wanting to install UMEM cards in some of my 
servers, from my analysis of the performance stats I expect that I could 
double the performance in normal operation with such hardware (cost is about 
$US800 for a 1G card).

http://www.coker.com.au/selinux/   My NSA Security Enhanced Linux packages
http://www.coker.com.au/bonnie++/  Bonnie++ hard drive benchmark
http://www.coker.com.au/postal/    Postal SMTP/POP benchmark
http://www.coker.com.au/~russell/  My home page

Reply to: