Re: tuning for large files in xfs
From: Nathan Scott
Date: Tue May 23 2006 - 22:23:26 EST
On Tue, May 23, 2006 at 06:41:36PM -0700, fitzboy wrote:
> I read online in multiple places that the largest allocation groups
> should get is 4g,
Thats not correct (for a few years now).
> I was also thinking that the more AGs the better since I do a lot of
> parallel reads/writes... granted it doesn't change the file system all
> that much (the file only grows or get existing blocks get modified), so
> I am not sure if the number of AGs matter, does it?
Yes, it can matter. For large extents like you have here, AGs
introduce a discontinuity that you'd otherwise not have.
> Sorry, I meant that moving the Inode size to 2k (over 256bytes) gave me
> a sizeable increase in performance... I assume that is because the
> extent map can be smaller now (since blocks are much larger, less blocks
> to keep track of). Of course, ideal would be to have InodeSize be large
> and blocksize to be 32k... but I hit the limits on both...
It means that more extents/btree records fit inline in the inode,
as theres more space available after the stat data. 2k is your
best choice for inode size, stick with that.
> > - Preallocate the space in the file - i.e. before running the
> > dd you can do an "xfs_io -c 'resvsp 0 2t' /mnt/array/disk1/xxx"
> > (preallocates 2 terabytes) and then overwrite that. Yhis will
> > give you an optimal layout.
> I tried this a couple of times, but it seemed to wedge the machine... I
> would do: 1) touch a file (just to create it), 2) do the above command
Oh, use the -f (create) option and you won't need a touch.
> which would then show effect in du, but the file size was still 0 3) I
> then opened that file (without O_TRUNC or O_APPEND) and started to write
> out to it. It would work fine for a few minutes but after about 5 or 7GB
> the machine would freeze... nothing in syslog, only a brief message on
> console about come cpu state being bad...
Hmm - I'd be interested to hear if that happens with a recent
> > - Your extent map is fairly large, the 2.6.17 kernel will have
> > some improvements in the way the memory management is done here
> > which may help you a bit too.
> we have plenty of memory on the machines, shouldn't be an issue... I am
> a little cautious about moving to a new kernel though...
Its not the amount of memory that was the issue here, its more the
way we were using it that was a problem for kernels of the vintage
you're using here. You will definately see better performance in
a 2.6.17 kernel with that large extent map.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/