🚀 go-pugleaf

In article <47ommk$d1v@nz12.rz.uni-karlsruhe.de>,
Arno Wagner <arno@ira.uka.de> wrote:
>alpha@america.net wrote:

>: I'm about to buy a new SCSI II hd, and have 
>: narrowed my choices to either a
>: Quantum Fireball, IBM Deskstar, or Conner CFP 1080
>: all are about the sam price, and same size, and
>[...]
>: Sustained data transfer is listed as 20 mb/sec
>: for the quantum and connor while the IBM
>: is listed as 4 mb/sec according to NecX direct.
>: So, does anyone have any experiance with any of 
>: these drives?  Assuming the above information
>: is correct, can you actually get 20 mb/sec?

>You can get 20 MBytes/sec but not in this price region.
>Sounds more like 20 MBit/sec = 2.5 MBytes/sec to me.
>The IBM has 3.2 MBytes/sec according to IBM doku, which
>can be found at http://www.ibm.com somewhere under
>storage / OEM storage.

To be completely fair to IBM's DeskstarXP, let me add:

That 3.2 MBytes/sec is a totally honest, ultra-conservative
worst-case figure for **sustained** transfer rate.  To wit :

  The real-data transfer rate on the innermost (ie slowest) tracks
  is 4.14 MB/sec, and since there are 4 data surfaces, a sustained
  transfer loses one disk revolution after every four revs, for
  the seek to the next cylinder.  Hence, 4.14 * 4 / 5 = 3.2 .

  [ Question for SCSI firmware hackers: Doesn't anybody optimize this
    situation by reading-to-cache (out of order) the data under the
    head, rather than waiting for the in-order data to come around ?? ]

I heartily commend IBM for putting forth such an honest # ,
even though it can be so easily misinterpreted, and lead
to a negative comparison relative to its competitors.

The non-sustained, real-data xfer rates for the DeskstarXP are:

Outer tracks:  5400 rev/min * 122 sectors/rev * 512 bytes/sector / 60 sec/min
             =  5.62 MB/sec  (M = 10**6)
Inner tracks:  as above, but 90 sectors/rev
             =  4.14 MB/sec

The Quantum Fireball 1080S, while also a 4-surface, 1.08GB drive,
has a different data layout. with fewer total cylinders, but more
sectors per track (rev); also, the # sectors/track difference
between outer and inner tracks is much greater: 88 inner vs.
177 outer.  So, the FB should be able to transfer (non-sustained,
outer-tracks) 8.15 MB/sec.  *But*, due to an apparent shortcoming
in its SCSI interface, it falls short of that, getting something
like 6 MB/sec on its outer tracks.

I would expect that the EIDE version of the FB might actually
get the 8 MB/sec (outer) and could be a best bet for audio-video
applications where sustained data xfers dominate the performance
choice (but you must be able to accept up to a 1:2 variation in xfer
rate depending on where the data sits).

Note that the avg seek time of the DeskstarXP (10.5ms) is
12.5% faster than for the Quantum FB (12ms), and this can
more than make up for the FB's slightly faster (real)
transfer rate in typical disk usage patterns.  Also,
the transfer rate of the IBM is more consistent across
the full surface, and it has a much larger *usable* cache
size (448k vs 80k).

I hope this analysis will encourage folks to do their
own "homework" and not rely on unreliable hearsay, and
fairy-tale MumbleMark "benchmarks".

In order for me to believe a disk performance benchmark, I
want its full premise defined, and results reported in MB/sec
xfers (of different xfer sizes), and average (over full
surface) seek times (that can be tricky for SCSI), *AND*
I'd like to be able to examine the source code of the test.

Short of that, I'll stick to the spec sheet from honest and
competent manufacturers.

-- brent