SUMMARY: seeking advice on 4/670 vs. sparcstation 10 model 52

From: James Ganong (jeg@diablo.ced.berkeley.edu)
Date: Tue Jul 07 1992 - 11:06:48 CDT


I wrote asking what are the benifits of a 600MP series machine vs.
a sparcstation 10.

i got mail from:

lm@slovax.Eng.Sun.COM (Larry McVoy)
guy@auspex.com (Guy Harris)
Jeff@digtype.airage.com (Jeff Wasilko)

also, i found out that sun has S.E.'s who are very willing
to talk about this kind of question. I talked to two people
in different parts of sun's S.E. organization:

Jim Seavey
Tony Vincent

Thanks you guys!

Briefly, everybody confirmed that for many purposes the SS10 will
be as good as a 600MP. The last note appended below has the most
technical info.

There are some advantages to the 600MP series:

The 600 series is available right now, without the supersparc chips,
and you can upgrade it when they come out. The multriprocessor
sparcstation 10's will probably be available in August(?).

I/O is probably faster on the 600 series, so if your job is
I/0 bound it might run faster on the 600 series.

The 600 series uses simms that are available from many sources,
but the SS10 will use its own simm design, which will take awhile
for third parties to clone.

The 600 series has vme slots.

For our site, vme slots don't matter, but availablity and i/o might.

Here are the replies:

From: lm@slovax.Eng.Sun.COM (Larry McVoy)

If you don't need the VME slots, just the CPU, buy the workstation. You
can only put a limited amount of mem (.5GB?) in a SS10, more in a 6xx.

--
---
Larry McVoy			(415) 336-7627			 lm@sun.com

From: Jeff@digtype.airage.com (Jeff Wasilko)

..his summary of my note deleted..

If you can, take a look at the 630. It's the same CPU board, just in a 5 slot deskside size. It leaves one 9U VME slot, and 3 or 4 SBUS slots (3 if you have two MBUS modules, 4 if you have 1 MBUS module).

The other difference is that the 630 can only go to 128 megs of RAM, vs 256 (or so) for the 670/690. One thing to check is that the SS10 may let you have 4 SBUS modules and 2 MBUS modules at the same time.

We've had a 4 processor 630 for about 3 weeks now, and I really like it.

From: guy@auspex.com (Guy Harris)

> it seems to me that a vme bus, and big box are not positive > features for this job, but i vaguely remember hearing that there are > bottlenecks (i/o? context switching?) in the sparcstations that have > been beefed up in the 600 series, but i don't know the specifics.

Probably I/O. From the standpoint of the CPU and MMU, there is *no difference* between a SS600 with a given set of Mbus modules, and an SS10 with the same set of Mbus modules.

> it looks like the ss6xx/41 is faster than the ss10/41. this supports > the 600-is-better-than-sparcstation theory i guess. from this chart > it looks like the ss600 and ss10 series are identical when you have > more than one processor, which would support the > identical-machine-in-a-different-box theory.

Umm, well, the "raw CPU" numbers are identical for an SS600 and an SS10 with the 40 MHz uniprocessor Viking module *and* the 50(?) MHz multiprocessor modules, which sounds as if it'd support a theory saying that the memory bandwidth over the Mbus to the CPU is the same on both the SS10 and the SS600.

The numbers that are better for the SS600 relative to the SS10 - even with the 40 MHz uniprocessor module - are numbers that relate to the *I/O* performance of the system, such as transactions-per-second and (for the uniprocessors) NFS ops/second. (NFS ops/second shouldn't improve much with multiple processors under SunOS 4.1.3, because there's a giant spin lock around the kernel, so that you can have only one processor running any significant amount of stuff in the kernel at a time; the NFS server runs entirely in the kernel. The multiprocessor versions of SunOS 5.0, which do much finer-grained locking, should have the potential of giving better NFS ops with multiple processors than with one processor.)

I.e., the SS10 and SS600 are identical *CPU modules* in different *systems*. The different systems seem to have memory subsystems of very similar performance, though (assuming the SS600 numbers aren't just the SS10 numbers with "SS10" crossed out and "SS600" written in in crayon).

> does the ss670MP have more cache memory than a ss10 with superCache?

No. The cache memory is on the Mbus module, so if they have the same Mbus modules, they have the same amount of cache. (As far as I know, the SS600 doesn't have any cache between the Mbus and the main memory subsystem; the cache all lives on the Mbus module.)

> does the ss670MP have more hardware page table slots?

No.

The only "hardware page table slots" to which you could be referring are TLB entries, and an SS600 with a given set of Mbus modules would have the exact same number of TLB entries as an SS10 with the same set of Mbus modules, because the TLB entries are on the Mbus module.

They also have the same number of contexts - 65,536 with the Viking modules.

No, that's not a typo. I really mean "65,536".

The context table exists entirely in memory.

In fact, the *entire page table* exists entirely in memory; all the Mbus-based SPARC machines (SS600 and SS10) have SPARC Reference MMU implementations, rather than the traditional static-RAM-based Sun MMU implementations.

> we are looking to get an smp machine and use it sort of as > a timesharing machine for the big modelling jobs, which would be > X applications displaying on the decstations.

To what extent is the work you're doing CPU-bound, and to what extent is it I/O-bound? If it's mainly CPU-bound, the SS10 should get you more bang per buck; if it's I/O-bound, the SS600 *might* get you better performance, but it might not.



This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:06:44 CDT