On 2 Aug 00, at 11:53, Jochem van Hal wrote:
> 2) Get a bunch of IDE-disk cabinets, and a lot of 70 Gb IDE disks,
> amounts to $25.000 or so (of course we will also need a small library
> for the copy pool). These cabinets have a scsi interface.
On 2 Aug 00, at 10:59, Roger C Cook wrote:
> ADSM considerations aside, from a hardware view in this environment
> SCSI is definitely the way to go. Not only is SCSI superior handling
> multiple simultaneous accesses, it is much less CPU intensive. When a
> SCSI drive is accessing, it normally uses 10-20% of the CPU, where IDE
> can use between 70-90%. That's why on a desktop system using IDE, when
> you transfer large amounts of data (such as copy one drive to another)
> the system becomes extremely sluggish during the transfer. SCSI will not
> normally have this impact. IDE may have evolved to surpass SCSI in some
> aspects, but SCSI is still the only choice in a server/multi-access
> environment.
>
This is not the point, because the RAID controller in these cabinets,
usually an i960 chip, handels the IDE load and presents the RAID set
as one SCSI (SE or LVD) disk to the server. We are using about 10 of
these boxes since two years for several purposes in multiuser
environments with good success. We never had any problem with them,
except the laughter from collegues ;-)
The throughput is not overwhelming, about 15 MByte/s, interestingly
regardless of RAID Level 0 or 5. There are new ones out with Intel
i960RN 64bit Processor which are said to be twice as fast for reading
and 50% faster for writing.
The point why we are not using these boxes but an expensive SCSI
RAID box as the primary pool with the ADSM Server is that in some
cases (archive, hsm) the version of some date on your primary pool
might be the only existing version. Given that
the memory cache of the IDE boxes consists of simple PC memory
modules without ECC
cache mirroring is not available
there is no meta communication between RAID and server.
there is no sniffing for bad blocks in the RAID software,
you cannot build several RAID sets with a global hot spare disk thus
having one 400 GByte disk which takes several hours to copy to tape
I would argue that this solution is - let's say one order of magnitude less
secure than a decent SCSI RAID. In other words, you have to decide
how much money the difference between 99,99 and 99,999 is worth.
However, if your choice is in fact, to have a primary pool on *tape*
without any disk pool (which makes sense only when you have very few
clients AFAIK) or to have one on IDE RAID's, I would recommend the
latter. At least, from my experience the ratio of unreadable tapes to
damaged disks to (usually due to operator mistakes) damaged RAID
sets is about 100:10:1.
Regards,
Michael Bruewer
----
Dr. Michael Br"uwer
Dr. Michael Br"uwer
RZ der Univ. Hohenheim 70593 Stuttgart
bruewer AT uni-hohenheim DOT de www.uni-hohenheim.de/~bruewer
Fon: +49-711-459-3838 Fax: -3449
PGP Public Key:
RSA: http://www.uni-hohenheim.de/~bruewer/pgpkey_V2
|
