[Lustre-discuss] Problems & partial success on PPC64 (XServe G5) Debian cluster

[Andreas Dilger]

On Nov 16, 2007  11:56 -0600, Robert Olson wrote:
>> As a starting point - we basically never test Lustre with a big-endian
>> server, so while it works in theory I would instead suggest starting
>> with a big-endian client and little-endian servers first, get that 
>> working,
>
> This works great - running a pair off OSTs on some intel boxes with clients 
> on the PPC64 nodes. Initial iozone measurements are making me happy, seeing 
> fairly decent performance over gigabit ethernet through at least a couple 
> switches (the servers are some older/slower machines that sit elsewhere in 
> the machine room from the cluster).

Good to hear.

>> and then tackle the big endian server separately (likely using something
>> like 2.6.22 ext4 as the starting point for ldiskfs, since the extent code
>> has proper endian swabbing already).  You could also try without mballoc
>> and extents on the OSTs.
>
> Are these changes that can be made by someone ignorant of the 
> implementation details of the lustre code? (config options, not applying 
> some patches, etc?) I'd be happy to try things out but would need something 
> of a roadmap to do so.

Well, you could start with the MDS on PPC, then try OSTs on PPC without
"-o mballoc,extents" mount options (you might need to pass
"-o nomballoc,noextents" to cancel out the former default options).

As for porting the ldiskfs patches to ext4...  I don't think it is necessarily
a simple task, but likely not impossible.  It should be pretty clear which
patches are already applied (extents, nlink, nanosecond), but porting some
of them (e.g. mballoc) would be very tricky (it is just about done in the
ext4 upstream repo) and at that point you can just mount without mballoc...

Cheers, Andreas
--
Andreas Dilger
Sr. Software Engineer, Lustre Group
Sun Microsystems of Canada, Inc.