Greetings,
About a week ago, I asked about people's experiences with mixing architectures
between servers and clients. In particular, we have a Sun4/280 with 6-4/110
dataless clients and may be adding Sparc2s. I also asked about the differences
between the hardware of Sun4/xxx and Sparc Stations, upgrade, etc.
I apologize for the tardiness of the summary, but email just kept on rolling
in. The following points are based upon the replies I received. Usually
I include the actual replies after my summary, because it lets people
draw their own conclusions. This time several people asked not to be
included in such a posting, so I'm not including the raw replies.
1) The difference between Sun4/xxx and Sparcs:
Same CPU chip ->> same instruction set, so applications are binary-com-
patible. The Sun4/xxx has a VME bus, the Sparcs have an S-bus. They
have different page sizes, different MMUs and floating point chip sets.
They are not compatible in anything that references /usr/kvm or when
programs are compiled with the -fast options.
2) Upgrade
Several people told me that Sun3s would be dropped from support, but
that was not the question. My users had the impression that Sun4s
which were not SparcStations also would not receive OS upgrades. That
doesn't seem to be the case, and the fact that the applications are
binary-compatible eases my mind somewhat about that question.
3) Mixing architectures
Everyone agreed that from the SysAdmin point of view, there was no
problem with having a Sun4/280 server with Sun4/110 and SS2 clients.
This mix is easier than a Sun3/Sun4 mix (although many people are
doing it) because there isn't as much space wasted in duplication
of applications -- all you need to duplicate is the kernel & /usr/kvm,
roughly 10 Mbytes.
One reply told me I'd have to use one of the SS2s to gen kernels
for the Sparcs and would have to fiddle with permissions on /sys, but
several others said that I could build both on the server if /sys/sun4c
was properly installed.
Capacity estimates I got ran 10-15 diskless clients, 20-30+ dataless
clients.
4) Caveats
The number of clients that the 4/280 can serve depends upon the amount
of memory in the server (and to a lesser degree, the clients). We have
16 Mbytes in clients, 128 Mbytes in the server, so that shouldn't be
a problem in our case.
The quality of the service depends upon the application. The 4/280 is
kind of slow to serve Sparc2s, but if they have their own boot/swap
disks it shouldn't be too bad. One reply cautioned against running
CAD applications on this setup, for instance, because of the large
size of the files being shipped.
Several users warned against mixing NFS service with compute service.
Unfortunately, practically all of our NFS servers (we have Sun3 and HP
clusters, too) are used for interactive service and background jobs.
This is something I inherited, and it is so ingrained here that the
situation probably won't change for some time. We do have a background
job policy which attempts to alleviate the situation. Some sun-managers
readers may remember my starting a discussion about this in comp.sys.admin
some time back.
I already had recommended to my user that he might want to get another
computer to use as a number-cruncher rather than the 4/280, because the
4/280 was not very fast (this is for a computer graphics/animation
research group). But while the machine isn't fast, it is underutilized.
That leaves us the problem of how to get more usage without overloading
the machine.
5) X-terminals
Responses to my question about adding X-terminals into the mix were
long on opinion and short on experience. Several people replied
"I've never done it, but it's no problem" or "I've never done it, but
it's a terrible idea". Others had experience but still were at
opposite poles in their answers, so I'm not sure what to say about
them.
My user wanted to provide a workstation for each faculty member and
supported graduate student at a new satellite facility. Since the
4/280 is underutilized, I thought that it might be a good idea to
hang some X-terminals off it to provide "seats" at lower cost. Then
more money would be available for a good compute server. I still
haven't made up my mind about this one.
6) Surprises
Apparently there are several products now that will boost NFS service.
One that was recommended was the Sun Prestoserve board. It is useful
when NFS write %age (nfsstat -s, then add write+create+remove) is
greater than 15%. It comes in VME-bus and S-bus versions. The board
is built by Legato. (request@legato.com with a subject of "index"
might get more info, but I haven't tried this yet) The person who
suggested this has been using it for about a month and is very satisfied.
Many thanks to those who replied.
--Sheryl Coppenger SEAS Computing Facility Staff sheryl@seas.gwu.edu The George Washington University (202) 994-6853
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:06:20 CDT