SPARCserver 390 user experiences - SUMMARY

[Philip Leverton]

Here is the summary of responses to my previous enquiry I made concerning
people's user experiences with SPARCserver 390s.

After I posted my initial query, SUN announced the SPARCserver 490 which
comes in the same enclosure as the 390 but has a new cpu/IO system.  A
concise description of the 490 (and other recently announced SUN products)
was posted to this group/list by Steve Grandi ("SUG trip report",
<4119 at brazos.Rice.edu>) -- thanks Steve.

Anyway here are the responses, my thanks to those involved. (A few
comments of mine are in brackets [])

Philip Leverton,
Computing Services, University of Melbourne
Australia

******
X-From: Donald Pace <pace at stat.fsu.edu>
X-Subject: Re: SPARCserver 390s - user experiences?
X-Organization: Dept. of Statistics, Florida State Univ.

We have had nothing but good experiences with our 4/390 once we got it.
It was ordered with a 90 day delivery promise and we did not get it until
150 days ( though it was one of the first shipped and I know that they now
are delivering on time ).  I am serving 5 diskless sparkstation 1s and 4
sun 3/50's with no observed problems.  We are also running compute bound
jobs on the machine, on average 2 at a time.  I believe that there are 16
contexts built into hardware.  In the past, we have only had problems with
contexts on out 3/160 ( 4 contexts ) when either 4 people were really
crunching or there were more than about 6 or so people reading their mail.
If the context question bothers you you might investigate the SUN 4/490.
It apparently has a new IO system that should fix some of these types of
problems ( according to the sales hype ).

******
X-From: raob at mullian.ee.mu.OZ.AU (Richard Oxbrow)
X-Subject: Re: SPARCserver 390s - user experiences?
X-Organization: Dept. of Electrical Engineering, University of Melbourne
******
[ description of DECR's (Department of Engineering Computer Resources) 390 ]

The configuration is 32MB Ram and 2 ethernets (ie0 and le2 ) and IPI
controller with a 1.2GB cdc saber attached. They are planning to use it to
server 20 3/50s ,plus provide application-disk service for another 30
3/50s !.. Only a few selected users will be allowed to login ..

Judging from mullians performance [see config below] a sun 4/390 should be
able to carry heaps more than 16 users with no problem. This year mullian
ran fairly smoothly with a max. of 27 users logged in and servering around
10-16 3/50-60 s all at once. My guess is that the amount of space
"swapped" out is the major factor in a SUN 3 or 4 in slowing down  , ie
this is explained by loggin into to munnari* [ a SUN 4] just after lunch
has or using a sun3/50.

mullians configuration: 

	sun 4/260 (no screen, you don't want a screen on a server)
	16 Megabytes of RAM (bare min. required !)
	SMD-4 controller (SMD-E)
	2*688 Fujitsu disks
	2*ie ethernet connections

richard ..

#ps munnari* after lunch munnari has alot of _rundaemons going which eats
up alot of swap space as well alot of other things.

******
X-Subject: Re: SPARCserver 390s - user experiences?
X-Organization: Deakin University
X-From:    Craig Warren <ccw at charlie.cc.deakin.OZ.AU>

We have had one on the floor for about 2 months now. In general terms, it
arrived, Sun came down to install it (which involved plugging it into the
power!), a couple of hours later we had SunOS 4.0.3 on it, and the
following weekend we turned it into a production machine.

Our 4/390 is a compute server if you like. We are running one of the
University's administration systems (student records) on it.  It averages
anywhere between 25-40 users. With our merger with Warrnambool Institute
of Advanced Education, we expect it to be running somewhere between 30-45
users on average. All these users are doing database type work. Mostly
reading with a little bit of updating.

We have 6 days of the year (enrolments), when it is just used for
updating. This is a very intensive period, and last year dragged a 4/280
to its knees. It will be interesting this year on the 4/390. All in all,
performance is good, but our users are on and off the system, so we might
only ever have 20 concurrent users at any one time.

>SUN advertising says that they can support up to 68 users and 15-30
>diskless NFS clients. But how many active users can they actually support?

We have certainly had up to 40 users on the machine. I suspect only about
20-25 were actually active. I would say it could easily support more
however.

>This is why I am interested in people's actual load they have on these
>machines. (I have heard that SUN4's get a slowdown after about 16 users'
>are active - I don't know whether I believe this or not - it seems
>implausible, and I hope very fictional for SPARCserver 390's.) For IPI
>disk/controllers, what sort of I/O rate can you hope to achieve (peak or
>averaged)?

IPI is a bit of a disappointment at this stage. We upgraded from the 4/280
to the 4/390 just so that we could get an improvment in disk write
performance. This was killing us at enrollment time.  Unfortunately the
disk provided on the IPI controllers are still only capable of 3Mb/s
transfer rates and 15 millisecond seek times. This means that whilst the
controller is a generation ahead of SMD,SCSI,etc, the disks are at this
stage, and are slowwing down the i/o system. Sun also have a problem with
disk write we believe, and we have been working with a group in the US to
help sort this out.

I would still feel reasonably happy about supporting 50-60 students on a
machine like this, as long as they weren't all running large lisp jobs or
something like that.

>Related to this is something I read recently in a UNIX mag that said that
>some implementations of SPARC have memory management units with eight
>contexts, and performance drops after more than eight jobs are being
>scheduled round-robin. Is this for real and exactly what SPARC
>implementations suffer from this? 

The new Sun 4/490 does not suffer from this, as it has more contexts. I
would benchmark the system before I considered this a problem.

******
X-From: guy at auspex.com (Guy Harris)
X-Subject: Re: SPARCserver 390s - user experiences?

>(I have heard that SUN4's get a slowdown after about 16 users'
>are active - I don't know whether I believe this or not - it seems
>implausible, and I hope very fictional for SPARCserver 390's.)

It depends on what "active" means.  See below....

>Related to this is something I read recently in a UNIX mag that said that
>some implementations of SPARC have memory management units with eight
>contexts, and performance drops after more than eight jobs are being
>scheduled round-robin. Is this for real

All but the number "eight" is real; I don't know how many contexts the
SPARCStation 1 has, but most other SPARC-based machines from Sun have 16
contexts (definitely including the 4/2xx and 4/1xx, probably including the
SS3xx).  (I think the SS 490 has 64 or so.)  Sun-3s other than the 3/80
and 3/4xx have 8 contexts (the 3/80 and 3/4xx use the 68030 on-chip MMU
and don't have "contexts" in the same sense).

If "active" (see above) means "running a program not blocked waiting for
the user to type something", then having more than 16 users "active" means
you have more than 16 jobs in the run queue, so that not all of the jobs
can have their address space set up in the MMU, and therefore when the
system switches contexts it may have to unload some old address space and
load up some new one, which does slow context switching down.