Hardware and benchmarks

[deraadt]

In article <1002 at brchh104.bnr.ca> netcom!tilman (Tilman Spokert):
   In article <988 at brchh104.bnr.ca> avri at asherah.clearpoint.com (Avri Doria):
   >I am looking for information which will help me choose between the Sun
   >4/4[7,9]0 and Sparc II as NFS servers.  I want a 470, but I have someone
   ...
   >Can a fully loaded server (i.e. maximum client population as defined
   >above) still be used as a compute server (i.e. cross compilation). As a
   >mail and news server?

   Don't do this. What gets easily overlooked is the fact that NFS is a set
   of user-level processes. So just one compile job would compete with NFS
   requests for system resources, like CPU, memory, IO, bringing down the
   overall server performance for the rest of the world!

I'd agree, but for different (more technical?) reasons.

My memory says the 4/4x0 has many more contexts in the MMU and caches. A
Sun MMU is really just a big funny static ram. So, on every context
switch, they would have to reload it for the new process. Same with the
cache. To avoid this, they put multiple contexts in their MMU and change a
little 3-6 bit value to pick which context. Then they roundrobin (or in
some other way) timeslice the contexts in the MMU and cache.

On a Sparcstation II, I think you get 16 contexts. On a sun4/4x0, I think
thats 32? Anyways, I'm pretty sure the 4/4x0 has more. The result is that
the mmu configs and caches have to be flushed less often, and as a result
more work gets done.

In general, as soon as you have as many simultaneously running processes
as the number of contexts, you are going to see performance degradation.

Now a question. Are current generation SCSI disks competitive in
performance with SMD disks?

Theo de Raadt
deraadt at cpsc.ucalgary.ca