[ILUG] Hardware and Software raid

[Paul Jakma]

On Thu, 11 Jul 2002, Ronan Waide wrote:

> On July 11, paulj at alphyra.ie said:
> > 
> > Most host CPUs will be much faster than the embedded CPUs found on 
> > hardware RAID cards. Important for RAID5.
> 
> Not to doubt you or anything, but this is the second time you've
> quoted this. I'm curious to know where it comes from.

cause i've never seen good hardware RAID5 write performance (which is
where the checksum overhead is) - /except/ for numbers Dave Rynne
posted from his ftp.esat.net machines. And he's got quite a tasty
setup, rather interestingly using Mylex ExtremeRAID - which have the
beefiest CPUs that i've heard of being used on RAID cards. (SA110 @
233MHz iirc). Also, iirc, his numbers were from 4 or 5 or more disk
RAID5 arrays. (and i cant find his email - but iirc he got over
100MB/s on reads, but still not over 25MBs on writes).

hang on, found the mail:

http://www.linux.ie/pipermail/ilug/2000-June/019174.html

his figures vary a bit - eg writes of:

Mylex ExtremeRAID:

	8x 7k: 7949
	5x 10k: 27738

MegaRAID:
	5x 7k: 15555
	3x 7K(?): 5393	

Perhaps differences in RAM and usage? (i suggest the lower numbers are 
the more representative numbers). The 8x case might be degraded due to 
bus contention though - it should be faster.

But I consistently see soft RAID5 beating hard RAID5. 

eg:

hardware:

              -------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
	  100  1387 87.8  6609 13.1  2670  7.2  1278 84.7 12569 27.0 581.6 10.8


software:

              -------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
          100  1408 90.1  9225 33.8  4229 30.6  1409 95.6 19223 66.8 1119.0 29.9


same machine, both 3 disk RAID5. Admittedly the RAID card is very old
(Mylex DAC960, i960CF (25MHz??), 8MB) so that goes against the
hardware raid figures, but then on the software RAID side the CPU is
an AMD K6 @ 200MHz. But then again, the hardware RAID setup has 2
channels - one disk on a seperate channel to the other two and further
that 1 disk is a 10k rpm disk (all UW scsi). for the software RAID
they're LVD - but they're all on the same channel.

NB: the file size is too low for the amount of RAM in the machine 
(100MB file size bs 128MB RAM). another bias. But it holds for both.

But then my 3 disks + sub EUR150 SCSI controller managed to beat 2 of
Dave's 5disk+ and EUR1k multi-channel hardware RAID controllers on
write performance...

but hey benchmarks.

Anyway, show me a hardware RAID setup that can get better write
performance than software RAID (with as few variables as possible).

And if you do, i'll point to the price tag. :)

Although, i reckon hardware RAID5 will win over software on read
performance in extreme cases.

(cause it'll come down to bus overhead and the RAID controller doesnt
have to send redundant data - plus the RAID controller is closer to
the disks and can discard reads of redundant data in certain
situations)

> After all, it seems to me that 99% of the point of having a hardware
> RAID card is so that you can just fire data at it and let it worry
> what happens that data w.r.t writing it to a disk array. Seems like
> this is a complete waste of time and money if the data's going to be
> RAID'd in the host CPU anyway.

I'm not suggesting use of software RAID on hardware RAID. I'm 
suggesting that in some cases you can save a reasonable amount of 
money, and get an /increase/ in performance.

i'm not biased towards either, i just think that the "only ever use 
hardware RAID" mantra that sometimes goes around amongst admins (with 
knowledgeable nods of agreement) is not completely justified.

Next contentious topic:

IDE sucks, but difference in suckage between IDE and SCSI is so small, 
and the price/MB difference is so great that SCSI is effectively dead.

> Waider.

regards,

--paulj