> Hi,
>
> Sorry to hear about all the layoffs. You guys have done a great job.
>
> I have been running XFS over RAID 5. However, I heard that Martin was
> still working on getting better performance out of XFS when RAID starts
> to run in degraded mode.
>
> Did I get my wires crossed and there was no such mini project?
>
> What is the status of XFS and RAID 5?
>
> Thanks for everything,
> Kal.
Well, there are probably things we can do in this direction, and Martin is
not totally gone, although he has been MIA for a few weeks now, I think he
is taking some time back in Denmark and is out of touch with email.
The key is to use buffer heads which cover a larger range if at all possible,
md and lvm both work by redirecting a buffer head to a new location, they
never split a buffer head into two. XFS has some 512 byte chunks of metadata
and the simplistic approach taken so far is that we do everything at this
size to avoid crossing boundaries with a buffer head. However, it would
be possible with some knowledge of the raid layout to say that a buffer
head woud not cross a disk boundary, and use larger ones. So for example
a striped volume with 4K boundaries would I think be safe to use a
single buffer head per page approach.
Not having spent too much time inside the md/lvm code, I do not know how
hard or easy working this out would be.
Steve
|