[ILUG] Hardware and Software raid
Paul Jakma
paulj at alphyra.ie
Fri Jul 12 04:47:49 IST 2002
On Thu, 11 Jul 2002, Ronan Waide wrote:
> Again, that's a "fact" without anything to back it. The RAID setups
> I've worked have been far more concerned with speed of access to
> data and use of logically large disks for things like databases.
hmm... some need speed, some need security. most need some kind of
compromise between the 2.
> Aside from which, the above still doesn't make sense: why is it
> safer to send your data to a raid controller only to have it sent
> back to be checksummed/replicated/whatever before beeing written
> down to disk? Seems like you're increasing the complexity of the
> write process, and more to the point increases the risk of putting
> inconsistent data on the disk.
you've misunderstood, the process is, more or less:
checksum data -> write to disk
for hardware raid the data goes:
ram -> pci bus -> raid controller (checksum/redundant data calculated)
-> disk
software:
ram -> cpu (checksum/redun. calc'ed) -> pci bus -> disk controler -> disk
ie the difference is in whether you get the host cpu to do the
checksumming calculations (at the cost of having to transfer more data
across the bus to the disk controller) versus sending just the data
across the bus and having a raid controller do the rest (but it might
have quite a scrawny cpu).
just to add some more numbers in the mix:
-------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
2000 10405 98.5 15143 9.6 3271 3.1 5708 60.2 9876 4.1 159.9 1.0
this is a 3x 80GB Maxtor software RAID5 setup with dual 1.something
Athlons. The machine might or might not have been busy during (parts
of) the above writing other stuff to disk: streams of jpegs,
additionally, it has a constant load on the PCI bus (regular transfer
of video data from 3 different cards - raw frames of PAL video @ bout
5fps).
15MB/s writes isnt bad under those circumstances... :)
(i'm speculating that perhaps read performance was penalised more by
the other loads on this machine than the writes were for some reason.
cause the read figure is quite low just under 10MB/s, and seeks are
abysmal.).
> I'm just looking for some backup (ha!) on these facts, I guess. Not
> doubting that you don't have it or questioning your expertise or
> anything.
well, those are just the opinions i've come to. be happy to be proven
wrong - great place to learn how wrong you are, this list is.
> Cheers,
> Waider.
--paulj
More information about the ILUG
mailing list