NT Ws supports striping, but I don't <g>.
It is possible to stripe as many drives as you can hang
on the box up to (nominally) 32, or at least according
to the documentation; I haven't tried it myself.
I don't have any info on performance of NT striping, nor
do I know the details of the implementation. Given
proper implementation, significant gains could be achieved,
possibly up to N times where N is the number of volumes
less some unknown (to me) overhead due to multiple I/O
operations and the inevitable competition for resources
at various levels (bus, controller, OS, memory, etc, etc).
Even though SCSI can handle multiple simultaneous outstanding
operations, the bus throughput is limited and there is
contention on it.
There is always some overhead. Often it is significant,
even when there is theoretically enough aggregate bandwidth.
There are plenty of shared resources in the chain with
the inevitable queueing around each resource. If you're
familiar with queueing models, you know that throughput
ALWAYS contains a factor of 1/(1-resourcebusyfraction).
Other factors can depend on scheduling algorithms, but
that one can't be eliminated. Practically speaking,
the slowdown becomes noticeable at around 10% busy. It
is EXTREMELY difficult to drive such resources beyond
about 70% busy, especially if there is a lot of variance
in service times (another factor is the mean square
deviation of the service time -- it's called the
dispersion for good reason).
I don't have any good feel for how this comes down for
striped disks. Clearly it depends on the controllers,
the distribution of disks, and the software implementation,
among many other things. One "other thing" be on-board disk
cache on the disk side of the controller bus (a tremendous
advantage because it sits at the other end of all the queued
shared resources and is used only by the single disk drive).
I don't know where a caching controller's cache sits relative
to the bus bottleneck, or even if there is a predefined place
for it.
Sorry not to be more help. To my mind, striping is just too
risky, though I can see that for work files, swap files, temp
files, etc, there could be major uses in certain environments.
I trust the disks a lot more than I trust NT, actually. I'd
be ESPECIALLY worried about NT's suffering a minor glitch
and forgetting the stripe configuration than the
physical disks themselves. Then, bye-bye data.
Spots |
