home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.arch:8252 comp.benchmarks:1184
- Path: sparky!uunet!paladin.american.edu!darwin.sura.net!mips!sdd.hp.com!caen!hellgate.utah.edu!dog.ee.lbl.gov!nosc!rigel!wolfgang
- From: wolfgang@code-541 (Lewis E. Wolfgang)
- Newsgroups: comp.arch,comp.benchmarks
- Subject: Re: Sun 600MP Benchmark Anomaly
- Message-ID: <5652@nosc.NOSC.MIL>
- Date: 24 Jul 92 00:51:18 GMT
- References: <19190001@hpcss01.cup.hp.com>
- Sender: nobody@nosc.NOSC.MIL
- Reply-To: wolfgang@code-541
- Organization: NCCOSC, NRaD Division
- Lines: 71
-
- In article 19190001@hpcss01.cup.hp.com, markw@hpcss01.cup.hp.com (Mark Williams) writes:
- >wolfgang wrote:
- >> There is a known problem with SunOs 4.1.2 and the MP machines
- >> that is fixed with Sun patch 100575-02. Symptoms included poor
- >> performance (worse than a 490) under certain conditions. One
- >> wonders if Workstation Labs had this patch installed.
- >
- >> We have a 690MP with 4 processors and are of the opinion that
- >> it is the best thing since cotton sheets! It is a win for us
- >> as a computational server, nfs server and three segment network
- >> router.
- >
- >Since the patch came out in May and the article was in the June issue,
- >it was probably done without the patch. My own guess is that the patch
- >was done in response to the article, but it's just a guess.
-
- I think the patch was available earlier, but I might be wrong. Some
- Sun people e-mailed me after my post and indicated that they did NOT
- have the patch, and also aparently didn't ask for any help from Sun.
-
- >
- >But the patch didn't really fix the anomaly. It helped two processor
- >performance, but did not change the fact that 4-processor system ran no
- >faster than a two processor system, at least in all the benchmarks I've
- >seen. The two processor systems ran only slightly faster than a system with
- >one processor stopped (20%). This is pre-Solaris 2.0. I don't know what
- ^^^^^^^
- Hm, how do you "stop" a processor?
-
- >2.0 will do, but I think software might not be the only culprit. MBus
- >may be saturated or the I/O system may be too slow. It's interesting that
- >future Sun Dragon servers do not seem to use MBus, according to published
- >reports.
- >
- >Bottom line: the 600MP (with SunOS 4 and the May patch) performs poorly as
- >an MP system, but it is a fine uniprocessor with better I/O than a workstation.
-
- I recall running benchmarks four times simultaneously (in background) and
- having each invocation perform only slightly slower than running it once on
- a 490. Running it once on the 690 yielded better results than a 490. I wish
- I had documented the execution times, but I was doing this only for my own curiosity.
- The system is in continuous heavy use right now, when things cool off a bit
- I will rerun the programs and document the performance.
-
- At the moment, the system has 5 users and two computationally intensive
- programs running all the time. At the same time it is serving 18 SparcStations
- and routing between 3 ethernet segments. I ran one benchmark that ran in 6.55
- seconds on the 690, 7.09 seconds on a lightly loaded 490. Another large
- acoustic modeling program just ran in 27 seconds on the 690, it runs in 28
- seconds on the 490. Looking at the perfmeter, idle times never got below 25%,
- indicating that it still had plenty of poop left. Hm, let me try running the big
- one twice in background (simultaneously), back in a moment... ok, I'm back. Each
- one ran in 36 seconds, versus 28 on the 490. Moreover, the system was still
- responsive, running 4 big jobs in background while serving and routing would really
- bog down a 490. Yup, it is a win for us.
-
- Could it be that you are thinking that one running job will utilize all available
- processors simultaneously? This system is coarse-grained, in that one process
- runs on only one processor. It may be swapped around to different processors, but
- is running only on one at any one instant in time.
-
- Luck
- Lewie
- wolfgang@nosc.mil
-
- NCCOSC, NRaD Division, Code 541
-
-
-
-
-
-