On Thursday July 27, [EMAIL PROTECTED] wrote:
> On Wed, Jul 26, 2000 at 08:12:47PM -0700, Gregory Leblanc wrote:
> > > Given the code at the moment, I am highly confident that linear, raid0
> > > and raid1 should be just as fast in 2.4 as in 2.2.
> > > There are some issues with raid5 that I am looking into. I don't
> > > know that they affect speed much, though they might.
> I doubt it, see below
I'm ready to be proved wrong ... you may well have suceeded :-)
>
> > Could you be a little more specific? Speed comparisons on disk access?
> > Then you can't compare RAID with no RAID effectively. You could compare the
> > speed of 2.2/2.4 RAID, and 2.2/2.4 no RAID, but comparisons across would
> > seem to be meaningless. Later,
> Ok, so here is what I did.
>
> Machine (you might have seen this before, sorry for repetitions)
> AMD Athlon 650, 256MB RAM, Abit KA7 Mainboard, VIA Chipset, system on
> Fujitsu 20GB IDE disk, 3 Promise PDC20262 UMDA Controller, 6 IBM
> DTLA-307045 46GB disks for data.
>
> It runs kernel 2.4.0-test1 (first 3 tests) and 2.4.0-test5-pre3 (rest)
> with Andre Hedricks IDE patch and the latest reiserfs.
>
> These are bonnie results:
>
> -------Sequential Output-------- ---Sequential Input-- --Random--
> -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
> Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
> fujitsu 500 2243 39.7 3400 9.4 1422 54.9 1945 75.4 2795 66.3 182.7 29.7
> single 500 5025 95.7 20218 81.1 11182 40.5 5214 90.9 26707 48.9 213.2 18.7
> raid0, 3 500 4804 83.3 17275 42.7 11067 50.7 5393 92.6 37242 65.2 361.4 28.1
> raid0, 4 500 4941 81.8 20920 48.7 13068 66.4 5283 91.0 35318 55.3 394.8 27.6
> raid5, 6 500 3009 52.0 3004 4.6 2589 3.6 4411 70.4 18982 13.9 360.5 22.7
>
> where
> - "fujitsu" is the system disk
> - "single" is a filesystem on a single IBM disk
> - "raid0, 3" is a RAID0 array of 3 IBM disks
> - "raid0, 4" is a RAID0 array of 4 IBM disks
> - "raid5, 6" is a RAID5 array of 6 IBM disks
>
> In between the disks have been freshly partitioned (1 big primary only)
> mkraid'ed and given a fresh reiserfs.
>
> It could of course be a combination of reiserfs, Andre Hedrick's IDE patches
> and raid. I expected the raid 0 with n disks to get a bit less than
> n times the block read performance of 1 disk, and raid5 to have block write
> performance of a bit less than single disk, block read performance much
> better than for a single disk. Are these expectations unrealistic?
>
I think your expectations of raid0 may be a little over optimistic.
raid0 will only get close to 'n' times a single disc when you have a
number of separate threads accessing the device, otherwise there are
fewer opportunities for multiple drives to be accessed at once.
I believe that bonnie is single-threaded, so it is unlikely to drive a
raid0 array optimally.
Possibly you could try running 3 bonnies in parallel, and compare that
to three paralel bonnies on three separate drives.
So your read timings look believable - raid0 is faster but not
stunningly faster. The extra speed probably comes from read-ahead,
which adds an element of multi-threading.
Your write times are a bit dissapointing though. The write-behind
that the kernel does should provide plenty of opportunity to get all
the drives writing at once. Cluster size can affect this a bit, but
I wouldn't expect it that much. I might do some similar tests my
self...
For raid-5, your numbers are disappointing, but not actually very
surprising now that I think about it.
The way the raid5 code is at the moment, only one request can be
outstanding for each 4k wide stripe. Also, request are processed
immediately, but are queue, and have to wait for the raid5d thread to
wake-up, which is probably a context-switch away at least.
I hope to fix this one day, but whether it will be before 2.4.0-final
comes out or not, I don't know.
Thanks for the numbers.
NeilBrown
