xfs
[Top] [All Lists]

Re: Linux XFS write performance

To: delusion@xxxxxxxxxxxx
Subject: Re: Linux XFS write performance
From: Chris Wedgwood <cw@xxxxxxxx>
Date: Sun, 17 Apr 2005 16:29:39 -0700
Cc: Michael Loftis <mloftis@xxxxxxxxx>, linux-xfs@xxxxxxxxxxx
In-reply-to: <4958.67.49.24.45.1113779121.spork@webmail.delusion.com>
References: <2ca133d205041622106b61c016@mail.gmail.com> <A2122DBD91F7C8AF935D0615@[10.1.2.230]> <4958.67.49.24.45.1113779121.spork@webmail.delusion.com>
Sender: linux-xfs-bounce@xxxxxxxxxxx
On Sun, Apr 17, 2005 at 04:05:21PM -0700, delusion@xxxxxxxxxxxx wrote:

> The card is a PCI-X 133 although there is also a PCI-E 8x card.  The
> PCI-E card uses a PCI-X 133 bus on the card so there shouldn't be
> much of a difference. The card supposedly can do the speeds but I
> haven't been able to get

I'm dubious that you can get 600MB/s on a single inexpensive card
right now.  I would try it on a raw partition and see how it performs
for you to get some idea of the upper limit.

This is hardware RAID right?  What level?  RAID-5 is going to be quit
a bit slower.

> 14 x WD raptors (73GB), 3.2GHz dual EMT64, 2GB, Supermicro
> motherboard.

FWIW on lesser hardware I can write 400MB/s across two controllers
(each doing RAID-5, I assuming RAID-0 would be faster).

> This is essentially a box that needs to backup data coming in over
> infiniband. The data will be written in 16mb files.

Can you actually sink 600MB/s over IB and also push that out over the
bus to the controller?

> Thanks. Is opening files on the fly till a bit slow as it used to be
> on IRIX?

It's not been overly slow or problematic for me.  What problems are
you seeing?

> Sounds like preallocating and opening is the best way to go then? Is
> there a fast way to preallocate in XFS (my xfs experience dates back
> a few years on irix and I need to catch up on what's been done since
> then).

'man xfsctl' for details on preallocation.  AFAIK it's the same basic
interface as what IRIX has used for many years.

> I did try sunit=256 and swidth=256*14 (I'm not testing raid 5/6 yet)
> but it didn't seem to help much. I also tried setting both to zero
> but it still didn't help.

I assume this is RAID-0 then?  Are you sure 128K RAID chunks are
optimal?  When I was testing I was suprised to find that values over
32K actually were slower than smaller values.

RAID-5 & 6 are going to be a *lot* slower for you I suspect (I doubt
the card you are using has enough bandwidth or CPU to deal with the
speed you are after with RAID 5 or 6 --- what doe the specs claim?)

> It's a fairly straightforward case...  16MB files all written
> sequentially.

16MB files are pretty small for this sort of IO rate.  In fact, if
they are all 16MBis in size --- why use a filesystem at all?

> In this case, would it be better to use one thread for writing or
> multiple threads?

Some experimentation would probably be needed.  I've found a small
number of writing threads is faster than one but a larger number has
no gain or is slower.

It would be nice to try AIO + DIO but presently that doesn't work and
I've not really had a chance to revisit fixing that (since apparently
it will ge done eventually anyhow).


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