"Martin K. Petersen" schrieb:
>
> >>>>> "Steve" == Stephen Lord <lord@xxxxxxx> writes:
>
> >> The XFS FAQ says that XFS performs slightly worse than ext2 on Soft
> >> RAID1 and RAID5.
>
> Hrm, I wonder why it says RAID1. RAID1 doesn't have a stripe cache
> and consequently doesn't suffer.
>
> I think the RAID1 comment may be a leftover from when I switched MD
> from 1K to 512 byte resyncs and had problems with the throttling.
>
> In any case, I think we should remove RAID1 from the caveat.
I'd like so see a comment there that using an external log with RAID5 is
a solution.
>
> >> XFS on Hardware RAID5 w/o write caching : ~10 min XFS on Hardware
> >> RAID5 w write caching : ~13 min EXT3 on Hardware RAID5 w/o write
> >> caching : ~13 min XFS on Software RAID5 w/o write caching : ~42 min
> >> EXT3 on Software RAID5 w/o write caching : ~12 min XFS on Software
> >> RAID5 w/o write caching, logdev on SoftRAID1 on the same disks :
> >> ~10 min
>
> Simon:
>
> Hrm, nasty. Did you let the RAID5 device finish resync before use?
> With both XFS and the resync messing with the stripe cache,
> performance is bound to suffer bigtime.
The RAID5 was always synced when I ran the tests (those U160/10000k
disks are really fast - and small with 9G).
The interesting thing is that with a simple bonnie or handling big files
the Soft RAID5 is much faster on this machine than the Hardware RAID5.
Streaming Gigabytes from a file (or was it to a file) gave me 112M/sec
with Soft RAID5 and 58M/sec with Hardware RAID5.
>
> Steve> We have talked about adding some padding to the log, but it is
> Steve> an on disk format change, so not something to do lightly, if I
> Steve> find time I may do some experiments with it.
>
> Another option (which could potentially also help XFS on mainframes
> with 4K hardware sectors) is to have a slim layer do read-modify-write
> on fixed size chunks between the fs and RAID5.
>
|