🚀 go-pugleaf

In article <C1BFFz.EEr@news.iastate.edu> john@iastate.edu (John Hascall) writes:
>alan@nabeth.enet.dec.com (Alan Rollow) writes:
>}thomson@zarda.macc.wisc.edu (Don Thomson) writes:
>}>We are looking at upgrading our campus Usenet News server from a hopelessly
>}>overburdened DECstation 3100 to a DECstation 5000/240.  The biggest bottleneck
>}>for handling news has been disk access time, so we're looking at options that
>}>will allow us the highest disk throughput possible to a disk with a capacity
>}>of at least a gigabyte.  I'm looking for advice on an optimal configuration
>}>for these needs, including:
> ...
>}Most of the I/O was 8 KB, but the next biggest chunk was 1 KB.
>
>}An IPI subsystem will probably help, but it seems a bit overkill...
>
>Perhaps not.  A fairly recent Usenix Paper, "Why SCSI is better than IPI
>for NFS" (or something like that), held that IPI was really only a win
>for large sequential I/Os, for the NFS I/O mix (generally 8KB and all
>over the disk [sound familiar?]) SCSI was a win.
>
>John

Actually on Ultrix systems this is not true.  IPI will still win.  The
paper uses a couple of arguments which are either outdated, or don't
apply to these systems:

1) SCSI should do better for random I/O since it has a buffer and IPI
   doesn't.

   The newest IPI drives now have buffers just like SCSI drives (Seagate's
   Elite-3 IPI drive has a .5 MB buffer).  Even when using non-buffered
   drives on DECsystems, this argument still fails since the IPI
   **controllers** extensively buffer and cache the data.  The paper
   compared a situation where the IPI drive was connected to the system
   with a much less sophisticated IPI **adapter**.

2) SCSI optimizes multiple drives by placing a lot of intelligence on the
   drive.

   Again newer IPI drives also have this increased intelligence built in.
   Also, the IPI controllers for DECsystems provides a much higher level
   of optimization than the SCSI adapters.  The IPI controllers can do
   overlapped seeks, command queueing, command ordering, rotational latency
   optimization, and command consolidation - all controlled by an on-board
   microprocessor which is running at 10 Million instructions per second.

Now as Alan Rollow points out, this may well be overkill for the original
requirement.  But even for random I/O or NFS; on DECsystems, IPI is much
faster than SCSI.  BTW, all the above remains true for TURBOchannel based
Alpha systems (DEC 3000/500 and DEC 3000/400) running OSF/1.


--
Joe Nordman, V.P. of R&D 				joe@genroco.com 
GENROCO, Inc.