Improving XFS desktop performance?

Improving XFS desktop performance?

[Pedro Ribeiro]

Hi all,

I've been using XFS since 2007 and I although I'm quite happy with its
performance, I am always looking at ways to improve it.
My usage is a regular desktop machine, for software development and
web browsing.
The main performance bottleneck I have is LUKS :-) However, I really
like my data encrypted, so dissing it is not an option. Actually, I'm
the xfs improvements discussed below I was able to get the performance
hit to something like 10% - negligible for me.

So what I do is specify a 128m log on filesystem creation:
mkfs.xfs -l size=128m /dev/mapper/target

And on /etc/fstab mount it as;
/dev/mapper/target      noatime,logbufs=8,logbsize=256k

After the above, my delete speed improved drastically.

Searching around the net I was able to find that using "lazy-count=1"
on mkfs would give a performance increase - how much would that be,
enough for me to do a full backup, format with that option + 128m log
and then restore all again?
And anything else you recommend?

Thanks for your time,
Pedro

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: Improving XFS desktop performance?

[Eric Sandeen]

Pedro Ribeiro wrote:
> Hi all,
> 
> I've been using XFS since 2007 and I although I'm quite happy with its
> performance, I am always looking at ways to improve it.
> My usage is a regular desktop machine, for software development and
> web browsing.
> The main performance bottleneck I have is LUKS :-) However, I really
> like my data encrypted, so dissing it is not an option. Actually, I'm
> the xfs improvements discussed below I was able to get the performance
> hit to something like 10% - negligible for me.
> 
> So what I do is specify a 128m log on filesystem creation:
> mkfs.xfs -l size=128m /dev/mapper/target
> 
> And on /etc/fstab mount it as;
> /dev/mapper/target      noatime,logbufs=8,logbsize=256k

FWIW, 8 logbufs is default already.

> After the above, my delete speed improved drastically.
> 
> Searching around the net I was able to find that using "lazy-count=1"
> on mkfs would give a performance increase - how much would that be,
> enough for me to do a full backup, format with that option + 128m log
> and then restore all again?

You should be able to use xfs_admin to turn it on; see the man page
for that.

> And anything else you recommend?

In general, the defaults are optimal for generic use; unless you have
a specific use case that is suffering, it's hard to make recommendations
beyond the defaults....

...although we do keep talking about making lazy sb counters default.

-Eric

> Thanks for your time,
> Pedro
> 
> _______________________________________________
> xfs mailing list
> xfs@oss.sgi.com
> http://oss.sgi.com/mailman/listinfo/xfs
> 

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: Improving XFS desktop performance?

[Dave Chinner]

On Fri, Nov 20, 2009 at 02:05:01AM +0000, Pedro Ribeiro wrote:
> And on /etc/fstab mount it as;
> /dev/mapper/target      noatime,logbufs=8,logbsize=256k
> 
> After the above, my delete speed improved drastically.

That would be the logbsize option your specified - logbufs=9 is the
default.

> Searching around the net I was able to find that using "lazy-count=1"
> on mkfs would give a performance increase - how much would that be,
> enough for me to do a full backup, format with that option + 128m log
> and then restore all again?

It depends on your workload. If you are doing lots of stuff in
parallel, then it makes a big difference because it removes the
superblock as a single point of contention. It also has the effect
of reducing the latency of fsync() if allocation was required.

Neither will affect typical desktop workloads unless you are
doing multi-media work, in which case the latency reduction of
lazy-count=1 make a big difference.

> And anything else you recommend?

No, the defaults set most of the best options for performance.
We probably should set lazy-count=1 as the default now that most
distros have picked up kernels that support that option now.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs