[Beowulf] SATA vs SCSI drives

Jim Lux James.P.Lux at jpl.nasa.gov
Tue Oct 12 11:49:07 PDT 2004

At 11:46 AM 10/12/2004 -0400, Mark Hahn wrote:
> > Though slightly dated, I hope the attachment is helpful....btw....I didn't
> > do an exhaustive search, but found the 10K SATA drives only offered at
> > 72GB's and under. The higher cap drives are 7200RPM.
>that's correct.  but remember - RPM is mainly for latency, not bandwidth.
>if your workload is not incredibly seeky, then you don't want to pay
>for latency, since higher density leads to lower cost, bigger disks, higher
>bandwidth and slower seeks.
>in summary:
>         - meet your reliability requirements using raid.  it's insane
>         to think about relying on a single disk in any non-ephemeral
>         setting anyway.  raid lets you achieve pretty much any reliability
>         you want (as well as offering a broad spectrum of performance.)
>         - meet your seek-rate requirements using RPM.  I find very, very
>         few applications are really seek-limited - really it's only very
>         databases with uniform-random distribution of reads of tiny data
>         from monumentally large tables.  in particular, if there's any
>         data locality or reuse at all, spend money on RAM not RPM.
>         - for anything large, get MTBF specs for prospective disks.
>         this lets you calculate how often you'll be replacing hardware,
>         physically.  your raid has taken care of data robustness;
>         this is purely a maintenance issue.
>there's no dramatic difference in any of the families of disks available
>(well, avoid 1yr warranties, of course!).  consider, for instance, that
>you can easily build raids based on 300G SATA disks that have half as
>many moving parts as with 147G SCSI disks.  even if the MTBF's differ
>by 50% (guess 1.0 and 1.5 Mhours respectively) SATA is more reliabile.
>it'll probably also be 1/4 the price and sometimes actually faster.

Read those MTBF specs carefully... Typically they'll have some sort of 
usage tied to it (so many seeks per second, number of power up/power down 
cycles).  ALso check the temperature effects on MTBF.  It's not unheard of 
for mfrs to specify MTBF assuming a 20C drive temperature, which is 
unrealistically cold.  Typically MTBF halves for each 10C rise in 

All the big mfrs have fairly decent descriptions of how they rate MTBF for 
their various drive classes.  Note well that the assumptions of use for 
drives intended for, e.g. consumer PCs, are very different from those 
intended for server duty, and this is primarily determined by how they are 
positioned in the market.

James Lux, P.E.
Spacecraft Radio Frequency Subsystems
Flight Telecommunications Systems
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875

More information about the Beowulf mailing list