[lug] hard drive speeds

[Jeffrey B. Siegal]

Sean Reifschneider wrote:
> A lot of people are talking *INTERFACE* speeds, when the more important
> thing in most low-end systems is the actual mechanism performance.

That's true in low-end systems, but the original message was asking about very
fast (i.e. high end) drives.  Some of the fastest drives approach 40MB/sec in
burst transfer rate (a few may even exceed that--I haven't looked recent), which
makes Ultra2 a necessity to get maximum performance and, as I indicated, don't
put more than two drives on a single channel.

> I'll second the recommendation -- while I haven't done side-by-side
> tests, IBMs are the only drives I use these days.

I agree with this, though not for performance reasons (I have no data on that). 
I love IBM drives.  They are quiet, well-documented, and very reliable.

> Remember that while spindle speed is an easy measure to get, density is just
> as important.  I would expect a 5400RPM 18GB drive with two platters to give
> better performance than a 10kRPM 18GB drive with 4 or 5 platters.  If the
> density is twice as high, it effectively doubles the throughput.

The numbers to look at are spin rate, seek time, and burst transfer rates.  How
to interpret these numbers and make trade offs depends on the application.

> This is one place where IDE drives are really winning.  For some reason, the
> really high density drives tend to be coming out IDE first.  Modern IDE
> with a single drive per channel is quite a respectable performer.

In terms of price/performance, I would agree that IDE is respectable. However,
the highest performance drives are still the Ultra2 SCSI drives, but you have to
be willing to pay for it.

> As Rob mentioned the biggest win is if you can increase the number of
> mechanisms providing data.

This also depends on the application.

> Remember that with UltraSCSI, you have SERIOUS cabling limitations --
> 1 foot between devices, 1 to 1.5 meter max cable lengths limit you to
> around 3 devices.

Yes.  I would not recommend UltraSCSI at this point, due to cabling limitations
and also sensitivity to cabling problems.

> If you're using 10KRPM drives, you'd
> better make sure that your case can handle it.

The best thing to do with high speed drives, unless you have a server case you
know can handle it, is to put them into a high quality external enclosure with
plenty of cooling.

> >Usually the lower sector numbers (which usually correspond to the ouside edge of
> >the platter) are faster than the higher numbers (inside).  Partition the drive
> 
> Unfortunately, there's no real way to tell how a particular drive allocates
> it's sectors.

I said usually.  The first ZBR drives were all over the map but more recently
most drives have tended to allocate the lowest sector numbers to the outside of
the drive.  I agree that if you really want to be sure you need to test.

> There was a fascinating presentation at Usenix a couple of years ago where
> somone was trying to apply these sorts of optimizations to the file-system,
> and he found that the only way one could tell was to actually benchmark
> every drive mechanism.

A couple of years ago this was more true than it is today.