RE: PLEASE TEST: IPI deadlock avoidance patch

I wrote:

> Hmm, after running the test a second time, I actually got a 
> crc error even when using a UP kernel. The test set is 9GB+
> in ~270 files. With SMP I seem to get around 10% corrupt
> files after simply making a duplicate of the folder. With
> UP I have gotten 1 corrupt file in 2 tries (~550 files
> copied).
> 
> I'll report back here after I have tested "classic" vinum.

Ok, I've run a few more tests now, and the results are not looking very good
for gvinum in RAID-5 mode:

Testbed:
--------
* FreeBSD 6-CURRENT, 2004.08.29.21.00.00
* ASUS P2B-DS mobo, 2 x 700MHz P3, 1GB ECC RAM
* OS disc: Some old 9GB ATA connected to on-board Intel BX chipset
* Data discs: 4 x 120GB Maxtor SATA discs connected to a HighPoint
RocketRAID 1540 card (HPT374 chipset)
* RAID-5 array created using "vinum raid5" command (so disklabel was edited)

Testset:
--------
* 9.1GB data in 273 files (mostly zip-files, rar-files and exe-files)
* crc checksums created and checked with /usr/ports/security/cksfv
* cd /space/raid5/ && cp -Rp testdir testdir.copy

Results:
--------
gvinum / SMP: 117 corrupt files in 4 tests (~10.7% error)
gvinum / UP : 3 corrupt files in 5 tests (~0.2% error)
vinum / SMP : no corrupt files in 3 tests


Observation: vinum seems to offer more than 3x the read performance compared
to gvinum (measured by timing the crc-checksumming). Write performance seems
about equal.

/Daniel Eriksson