[Beowulf] GPFS on Linux (x86)

[Craig Tierney]

Mark Hahn wrote:
>> However, what if you know you need better access time or more reliable 
>> better disks (SCSI is still more reliable than SATA in most cases) and 
>> want to invest in 10k or 15k SCSI disks?
> 
> I'm not sure I've met anyone who really knew that rather than believing 
> it on faith ;)
> 
> for instance, it's not trivial to argue better SCSI access time in an 
> environment where you have many disks.  I suppose if you actually
> have a _single_ thread of IO, you really will always only care about
> the latency of that single outstanding op, but that seems pretty exotic.
> striping handles multiple streams pretty nicely (even striped journal
> devices, if you're doing massive updates...)

Fair comment.  I wouldn't be the one buying the 15K scsi disks unless
it was proven they are that much faster for my application.  I have 
heard some horror stories about SATA disk in raid systems and how
vendors have to work around quirks in the SATA disks to get them to
work ok.  They describe it as the SATA disks go 'out to lunch' 
occasionally and a disk will have a large pause before continuing. 
Those comments come from a vendor (Be wary).

But if my customer wants 15K disks, I am going to get them for him/her.
I am not going put the effort in convincing them to go with SATA.  You
pick your battles even when serving them.

> 
> reliability is even fuzzier because vendors are such jerks about 
> publishing their numbers.  in the broadest sense, the availability of 
> 5-year warantee,
> "raid-edition" SATA disks is kinda sorta equivalent to SCSI.  SCSI disks 
> do sometimes include MTBF specs (1.0-1.4 Mhours, iirc) - SATA is probably
> somewhere above 800K hours, though that's somewhat confounded by how you 
> figure the working load (24x7, 8x5, temperature, etc).  and the lack of 
> numbers from SATA vendors...

Large SATA array vs. Large SCSI array, same vendor.  It appears that the 
  SCSI disks were a bit more reliable.  I wish I had more numbers to 
back it up, but that is what it seemed like.  I don't buy any of the vendor
numbers.

Craig

> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit 
> http://www.beowulf.org/mailman/listinfo/beowulf