Re: [K12OSN] SCSI Disk subsystems was: Building a new lab server...recommendations?

["Terrell Prude', Jr." <microman cmosnetworks com>]

For what it's worth:

I used to do software RAID on NT boxes, both RAID 1 and RAID 5, with 
SCSI disks on a single channel.  I also tried RAID 1 with two SCSI 
channels and got virtually the same results.  The RAID 1 performance 
was, as I recall, about the same for reads, but about a quarter to a 
third slower for writes.  However, I could survive a disk failing.  RAID 
5 was about twice as quick as a single drive if I used four to five 
disks, and if the CPU was fast enough to do the parity calculations.

Then...I discovered hardware RAID 5, specifically, on Compaq servers 
(Pentium III's) with SmartArray 3200 and 4200 controllers.  What I saw 
was, depending on the application, apparent disk accesses (reads and 
writes) anywhere between five and fifteen times as fast as with a single 
SCSI drive (all were 18GB, 10k RPM Ultra2-SCSI disks).  This was thanks 
to the hardware RAID controller doing the parity, as well as the cache 
on the controller (64MB on the 3200's, 128MB on the 4200's).  Heck, it 
made even Windows NT look fast!  :-)

My experience with multiple disks on a single SCSI channel is unlike 
that of my experience with multiple drives on the IDE channel.  SCSI 
just handles this better, by design, than IDE, especially with 
bus-mastering controllers (e. g. anything from the Adaptec 2940 on).  It 
was built to multiplex, which is exactly why we have always used it for 
hardware RAID.  Let say you have Ultra-Wide (40MB/sec) SCSI.  We all 
know that these ratings of "30MB/sec" for disk drives are bogus in the 
real world; on real file systems (single drive, not hardware RAIDed), 
I've measured--tops--about 5MB/sec, and it's often in the hundreds of 
kilobytes range, due to the mechanics of the read/write head actuator.  
So, let's say that real-world performance for a decent SCSI drive is, on 
average, 3MB/sec.  Ultra-Wide SCSI can theoretically support up to 15 
devices on the channel.  Assuming you actually do fill up the channel 
like that, this makes 3MB/sec times 15 disk drives = 45MB/sec on the 
SCSI channel.  This means that if you lower it to 13 disks, you're 
filling up 39MB/sec (theoretical; actual will be slightly less due to 
the small overhead of having to talk to each disk), which is within 
Ultra-Wide SCSI's range.  We ran 14-disk Ultra2-SCSI arrays per channel, 
and it was suh-weet.

The moral:  Be less concerned about the speed of the SCSI channel and 
more concerned about the performance of the disks you're going to put 
onto that channel.

--TP

Steve Wright wrote:

> On Tue, 2003-04-01 at 01:17, G.Chester wrote:
>  
>
> > [....]  As a relative newbie I found various HOWTOs and review articles that
> > should go a large way toward satisfying the need called for in this
> > discussion thread.  Here is a starting list of articles that have helped me
> > understand a bit of this thing called Linux when it comes to RAID, hard
> > discs and file systems:
> >
> >    
>
> thank you.  There /are/ plenty of HOWTOs, and I was not thinking of
> duplicating those.
>
> The topic that does not seem to be addressed, is "WHY we would want to
> use one of these technologies over another."
>
> One of the delights of Linux (all *nix really) is that the filesystems
> are modular, and we may assemble them in the order that suits us, and
> for whatever reason that we choose.
>
> The HOWTOs, and general consensus, suggest to us that RAID is a panacea,
> and in the same handful, it will be FASTER !  I *suspect* this is not
> the case.
>
> With IDE for example, for outright speed we might decide to stripe two
> drives and then place them on the same IDE channel.  The theory is solid
> until we really think about it, hdparm finally revealing that our
> assumptions have been incorrect.  The IDE people have known that we
> should never place two drives on the same cable, no matter what.
>
> SCSI being quite a different animal, I do not know how many drives I can
> add to a single SCSI channel before that channel (assembled as a RAID
> device) will now be slower than a single drive, and maybe with the newer
> SCSI 320 controllers, that ceiling will much higher.
>
> At this stage it is only a hunch Igor, but I am willing to bet..
>
> The unanswered questions are ;
>
> 1.)  Striping for speed, are two RAID striped drives faster than a
> single drive of the same type WHEN THEY ARE ON THE SAME SCSI CHANNEL.
>
> 2.)  For redundancy, which RAIDlevel do I use, and what is the
> performance hit to place it all on one channel.
>
> Here's one test.  This is SOFTWARE RAID, a simple mirror, on a single
> SCSI channel.  The test runs on the md0 RAID device, and then on one of
> the component drives in that array (2 drives.)  This obviously not a
> modern machine (most of my stuff is junk.)
>
> /steve
>
> [root computech root]# hdparm -tT /dev/md0
>
> /dev/md0:
> Timing buffer-cache reads:   128 MB in  1.04 seconds =122.96 MB/sec
> Timing buffered disk reads:  64 MB in  3.38 seconds = 18.93 MB/sec
> [root computech root]# hdparm -tT /dev/sdd1
>
> /dev/sdd1:
> Timing buffer-cache reads:   128 MB in  1.04 seconds =122.96 MB/sec
> Timing buffered disk reads:  64 MB in  3.39 seconds = 18.86 MB/sec
> [root computech root]# lsraid -d /dev/sdd1
> [dev   9,   0] /dev/md0         96049789.24709966.10FBC473.A100194C
> online
> [dev   8,  33] /dev/sdc1        96049789.24709966.10FBC473.A100194C good
> [dev   8,  49] /dev/sdd1        96049789.24709966.10FBC473.A100194C good
>
> [root computech root]#
>
>
>
>
>
> _______________________________________________
> K12OSN mailing list
> K12OSN redhat com
> https://listman.redhat.com/mailman/listinfo/k12osn
> For more info see <http://www.k12os.org>
>
>