Newsgroups: comp.sys.alliant
Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!ira.uka.de!fauern!NewsServ!ries
From: ries@Informatik.TU-Muenchen.DE (Bernhard Ries)
Subject: Experiences with Alliant FX/2800
Message-ID: <1991Jun21.104122.9132@Informatik.TU-Muenchen.DE>
Sender: news@Informatik.TU-Muenchen.DE
Organization: Technische Universitaet Muenchen, Germany
Date: Fri, 21 Jun 1991 10:41:22 GMT
Lines: 52


We have an Alliant FX/2800 which has been running for about 6 months.
The configuration is as follows:

    Memory Size   = 64 Mb
    CE-Cache Size = 1024 Kb
    User Memory   = 48.9 Mb
    Paging File   = 165.3 Mb
    Swapping File = 135.3 Mb
    Number of IPs =  1
    Number of CEs =  8
	Detach/Attach = 5/4
	    Cluster 0: CE00 CE01 CE04 CE05 

We are currently running Concentrix 2800 2.0.04 and use the
following compilers:
    FX/C-2800 V1.1.27
    fortran , version 1.1.27

In the last couple of months we have experienced a number of
rather nasty problems with the machine:

1) Programs written using the threads package often cause 
   operating system panics. In one case, a C-program which
   used a wrong array index caused a double bus error. Another
   example recursively created threads. This also led to a
   operating system crash after a while.

2) The compilers (especially the FX/C-Compiler) often crash and 
   produce core-dumps. In other cases, loop unrolling is not
   handled correctly, which leads to completely wrong results
   (A simple loop that should print 1,2,3 prints 1,1,2 instead, etc.)
   Since this form of loop unrolling can not be disabled individually,
   the only way to fix this seems to be to use no optimization
   at all.

3) Excessive ethernet-traffic (e.g. starting a big number of
   applications that produce a lot of X-Window output) cause
   a deadlock. We circumvented this problem by doubling the
   buffer space allocated for the ethernet.

4) The speedup which can be achieved with the machine are
   disappointing. This applies to applications which use
   processor clusters and to explicitly parallel programs which
   use threads. One of the main reasons for this seems to
   be the lack of cache consistency mechanisms for the i860 caches.
   In most cases, programs which use the threads package are
   considerably slower than equivalent programs using standard
   UNIX processes.

I would like to know if anyone out there has an FX/2800 and has
made similar experiences or if we are just doing everything wrong.
