Re: ciss(4): speed degradation for Compaq Smart Array [edited]

Doug White <dwhite_at_gumbysoft.com> wrote:

> You've still omitted the array setup, including RAID type and stripe size.

Yes, sorry. I missed it in 2nd letter, but said at the first place,
that disk systems configured identically as RAID5 (5 disks in array).

Standard HP array BIOS has no nuts and bolts to tune array parameters,
so the arrays were configured with default options. As I remember,
there should be 64K default stripe size, but I'm not quite sure --
to find it there is needed additional array configuration utility.

do:~ $ grep da0 /var/run/dmesg.boot
da0 at ciss0 bus 0 target 0 lun 0
da0: <COMPAQ RAID 5  VOLUME OK> Fixed Direct Access SCSI-0 device
da0: 135.168MB/s transfers                                       
da0: 69443MB (142220640 512 byte sectors: 255H 32S/T 17429C)

re:~ $ grep da0 /var/run/dmesg.boot
da0 at ciss0 bus 0 target 0 lun 0                                
da0: <COMPAQ RAID 5  VOLUME OK> Fixed Direct Access SCSI-0 device
da0: 135.168MB/s transfers                                       
da0: 347295MB (711261810 512 byte sectors: 255H 63S/T 44274C)

do-test:~ $ grep da0 /var/run/dmesg.boot
da0 at ciss0 bus 0 target 0 lun 0                                
da0: <COMPAQ RAID 5  VOLUME OK> Fixed Direct Access SCSI-0 device
da0: 135.168MB/s transfers                                       
da0: 138911MB (284490240 512 byte sectors: 255H 32S/T 34864C)

> I'd also suggest using a tool like iozone to run your tests instead of dd.
> Unless your workload consists of entirely sequential writes this perf test
> is worthless.

Yes, perhaps, more thorough testing is needed, indeed.
While, I think, if you have bad performance with linear transfer,
all other tests would be worthless.
My old system had linear read transfer near theoretical controller bus
limit, while it's 2 and more times less with new.

And I hadn't intended to do "parrot" measurements.
Just, while upgrading my network statistics server, which should process
Cisco flows, I've noticed substantial performance drop.
Linear transfer is just an ilustrative example. It can use block sizes
of 1k, 64k or 1m, but the result is the same.
Same with other non-linear things, like simple tar of /usr/src.

Well, if it couldn't be fixed, perhaps, I should think of downgrading
system back to 4.11.


-- 
Regards,
 Andrey