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 19:17:43 -0700
Cc: Michael Loftis <mloftis@xxxxxxxxx>, linux-xfs@xxxxxxxxxxx
In-reply-to: <1308.67.49.24.45.1113783605.spork@xxxxxxxxxxxxxxxxxxxx>
References: <2ca133d205041622106b61c016@xxxxxxxxxxxxxx> <A2122DBD91F7C8AF935D0615@[10.1.2.230]> <4958.67.49.24.45.1113779121.spork@xxxxxxxxxxxxxxxxxxxx> <20050417232939.GB12606@xxxxxxxxxxxxxxxxxxxxx> <1308.67.49.24.45.1113783605.spork@xxxxxxxxxxxxxxxxxxxx>
Sender: linux-xfs-bounce@xxxxxxxxxxx
On Sun, Apr 17, 2005 at 05:20:05PM -0700, delusion@xxxxxxxxxxxx wrote:

> The 8port card seems to do fairly well but it remains to be seen if
> the 16port card can perform.

How did you test the 8-port card?  What results did you get?

> The claim it can but not many people have used the card yet.

I couldn't find details on what the manufacturer claims the card can
do.  Do you have the URL?

> Raid 0 at first and then hopefully raid 5 or 6 (dual parity) at some
> point with more drives/cards.

You realize RAID-5 is going to be slower for writes in many cases?
(Large writes won't be so bad, but small writes can hurt).  RAID-6
slower still.

> Which cards are you using?

I'm loathed to make recommendations on this hardware at present.  But
I think to get 200MB/s per card you can find several options.

> Mellanox claims they can get over 500MB/s with their SRP driver.

Sure.  But to do something with that data you need to move it about
twice or more usually.

> 500MB/s if fine for me but I'd like to have the additional disk
> bandwidth.

So you suck in 500MB/s into main memory --- and write that out.
That's 1GB/s of IO without any overhead.  Can you realiably do that on
your system?

> I assumed 128 would be better but I'll try 32 and 64k next.

For large writes it doesn't make much difference, but for smaller
writes larger chunk sizes hurt more.

You usually don't actually have that much control over the size of the
writes that will be sent to the card from userspace (clearly you can
influence this, but generally speaking writes will be fragmented and
even at times reordered underneath you, and there is also the
filesystem metadata updates to consider which are going to be small
even for large writes).


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