Re: [tlug] SATA software RAID or SAS hardware RAID?

[Curt Sampson <cjs@example.com>]

Well, mostly nitpicking here, except for the suggestion at the end:

On Thu, 14 Dec 2006, Jonathan Byrne wrote:

> On Saturday 09 December 2006 23:49, Curt Sampson wrote:
>
> Well, maybe one: SCSI disks are still more reliable than [PS]ATA disks,
> although those have gotten so reliable that this my no longer be an
> issue for most people.

If you look at a single drive, sure. But a 300 GB SCSI drive is about
US$600 these days, according to a quick check on the net. A pair of 300
GB IDE drives will set you back $300. Mirror the two IDE drives. Even
ignoring that the IDE drives are half the price for the same amount of
storage and probably better speed (depending on the application and your
OS), which system is going to offer better reliability?

> Another is that you can put several SCSI devices on a cable and still
> get good performance b/c the SCSI host adapter offloads the work from
> the CPU.

I'm curious as to exactly what work you think it's offloading from the
CPU.

It is true that one of the advantages of a single SCSI bus is that if
you put multiple drives on it, the ability to have multiple outstanding
requests amongst multiple drives at once is an advantage. But this
advantage evaporates once you actually look at a whole system, where if
you have multiple drives, you can also have multiple buses.

The only reason to put multiple drives on a single bus is to save the
expense of another controller. But IDE controllers are so much cheaper
than SCSI controllers that this "advantage" is entirely useless. Not to
mention that, with modern 7200 RPM drive transfer rates, any application
that's not entirely seek bound can't put more than two devices on a SCSI
bus anyway.

> Hardware RAID is still going to offer better performance than software
> RAID....

I disagree that this is even a general rule of thumb any more even when
cost is not taken into consideration. Various systems factors could
easily make software RAID faster than hardware in some circumstances.
Some servers, for example, have more than one PCI bus, spreading the
disk I/O load across the two buses with IDE controllers may well achieve
better performance than the maximum theoretical performance possible
with a single controller of any kind.

> With software RAID, you don't have this problem, and you also get a
> performance upgrade anytime you upgrade your CPU.

I'm willing to bow to benchmark data, if you have it, but I'll posit
from my experience and analysis that CPUs are so fast these days that,
except on extremely CPU-bound systems (where faster RAID is not likely
to help overall performance anyway) you will very rarely, if ever, see
faster disk I/O from upgrading your CPU. The processing power needed for
RAID is just such a small fraction of what a modern CPU can do that it's
pretty near inconsequential.

cjs
--
Curt Sampson       <cjs@example.com>        +81 90 7737 2974