shoebox vs ethernet swap on Sparcstation

[khb at sun.com]

In article <3260 at brazos.Rice.edu> it is written:
>X-Sun-Spots-Digest: Volume 8, Issue 205, message 17 of 19
>
>We are planning to get some sparcstation 1's for our system (280 server
>with 753 controller and 5 fujitsu smd disks, 16 3/50 or 3/60 clients with
>8 meg memory) and I am wondering if the performance of a local scsi disk
>(for swap only) is better than swapping over the network.  I will
>summarize if there is interest.
>

Note that I am a compiler weenie, not an I/O or OS expert ....

There is no definitive answer possible. It all depends on how busy your
network is, what disk your server has, and how much RAM the server has
(and % writes vs. reads) etc.

I had one code to run for a customer which was about 80K line f77 compile
to be compiled at -O4. The best time I was able to obtain on a 4/3xx was
with the swap and compiler temps on the net, served by a 4/280 with 128Mb
RAM and SMD drives. With the IPI drives on the server, performance might
have been a bit better (but I doubt it, as 128Mb was enough to reduce the
amount of disk traffic considerably).

Despite a background in estimation theory (kalman filtering), I can't come
up with a good mathematical description of _all_ the variables, much less
a closed form analytical model. Some key variables are:

	1) speed of server's CPU
	2)                   I/O
	3) server's load (CPU and I/O contention)
	4) network load/transport/effective bandwidth
	5) application's access pattern

There seem to be some variables missing (to explain what I've seen in some
experiments).

My advice is to figure out what your usual workload is, and to try both
configurations with a benchmark which resembles your workload.

Since OS 4.x's pages utililies from their a.out's rather than loading
followed by swapping out, so changing the performance of paging doesn't
always produce application performance results in the way one would
expect. The virtue of the ULTRIX style is that it is easier to predict,
and then a local disk is really mandatory (and that makes the bottom line
profit better). The 4.x style is certainly more entertaining, and the OS
folks are sure it really does go faster on average. 

If you come up with what purport to be definitive results, please post them.

Cheers

Keith H. Bierman    |*My thoughts are my own. !! kbierman at sun.com
It's Not My Fault   |	MTS --Only my work belongs to Sun* 
I Voted for Bill &  | Advanced Languages/Floating Point Group            
Opus                | "When the going gets Weird .. the Weird turn PRO"