building a RAID system
becker at scyld.com
Wed Oct 8 22:58:17 EDT 2003
On Wed, 8 Oct 2003, Mark Hahn wrote:
> > - get those drives w/ 8MB buffer disk cache
> what reason do you have to regard 8M as other than a useless
> marketing feature? I mean, the kenel has a cache that's 100x
> bigger, and a lot faster.
The larger cache does provide some benefit. Disks now read and cache up
to a whole track/cylinder at once, starting from when the head settles
from a seek up to when the desired sector is read. You can't do that
type of caching in the kernel.
As disks become more dense, more memory is needed to save a cylinder's
worth of data, so we should expect the cache size to increase.
But you point is likely "disk cache is mostly legacy superstition".
MS-Windows 98 and earlier had such horrible caching behavior that a few MB
of on-disk cache could triple the performance. This was also why
MS-Windows would run much faster under Linux+VMWare than on the raw
> > - it supposedly can sustain 133MB/sec transfers
Normal disks top out at 70MB/sec read, 50MB/sec write on the outer
tracks. These numbers drop significantly on the inner tracks.
You might get 10MB/sec better with 10K or 15K RPM SCSI drives, but it's
certainly not linear with the speed.
BTW, 2.5" laptop drives are _far_ worse.
Typical for a modern fast drive is 20MB/sec read and 10MB/sec write.
Older drivers were worse.
> > - some say scsi disks are faster ...
> usually lower-latency, often not higher bandwidth. interestingly,
> ide disks usually fall off to about half peak bandwidth on inner
> tracks. scsi disks fall off too, but usually less so - they
> don't push capacity quite as hard.
Look at the shape of the transfer performance curve -- the shape is
sometimes the same as the similar IDE drive, but sometimes has a much
different curve. Wider tracks mean faster seek settling but lower
Donald Becker becker at scyld.com
Scyld Computing Corporation http://www.scyld.com
914 Bay Ridge Road, Suite 220 Scyld Beowulf cluster system
Annapolis MD 21403 410-990-9993
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf