🚀 go-pugleaf

> Delphi Mac Digest          Sunday, 27 Apr 1986      Volume 2 : Issue 16
> 
> From: BRECHER (7507)
> Subject: Interleave/DiskBench
> Date: 22-APR 10:19 MUGS Online
>  
> Which Interleave is best?
> -------------------------
>  [results from SSDiskBench, which does single-sector operations]
> These results clearly confirm Bass's "about 6:1" estimate for optimal
> handling of consecutive single-sector requests.

Not really.  It actually confirms Bass's claim *assuming consecutive
requests are issued with minimal processing between them*.  The diskbench
program issued consecutive reads/writes as quickly as possible.  Do
most applications that do single-sector reads/writes do that, or do
they process each sector of data between operations?

> I gathered some statistics on the distribution of disk I/O requests by
> writing a hook into _Read and _Write which records the size of each
> request and the the sector (logical block) distance from the end of
> the previous request to the start of the current request.

The missing item here is the *interval* between requests.  Unfortunately,
the Mac's clock is to coarse to be much help here.  If the apparent time
between two requests is 0 or 1 ticks, they *might* be "back-to-back."
Or, enough time might have elapsed to defeat any attempted optimization
of the disk interleave.  If more than 1 tick has elapsed, you can be sure
that the disk interleave doesn't matter anymore, since typical disk
rotation times are on the order of 1 tick.

> So, whether 1:1 or 6:1/7:1 is better from the standpoint of getting the user
> home quicker depends on the user's job mix.  To find out what kind of
> activities favor interleaving or lack thereof, I wrote another _Read/_Write
> hook.  Here, if the request started in the same cylinder in which the previous
> request ended, and if there were five or fewer intervening logical sectors,
> the relative advantage of 6:1 vs. 1:1 interleave was calculated and
> cumulated.  And, the relative disadvantage of 6:1 in terms of sectors passing
> under the heads during actual data transfer was calculated and cumulated.
> The following is a general summary of the net winner by activity:
>  
Again, all this assumes that single-sector requests are issued with
no intervening processing.  I suspect that the real picture is very
different.

Ephraim Vishniac
decvax!wanginst!wang!ephraim