Here is something interesting. We got an OS/2 SMP machine in for
evaluation (2 90Mhz Pentiums), and it was running our software very
slowly (I mean VERY slowly.) At first I thought, "Oh no, we obviously
have some concurrency problems that were not evident on a uniprocessor
system." So I wrote a quick test program that showed up some
interesting issues.

The program starts up 1 to 5 copies of a thead, which just does some
CPU intensive busy work. If we start both processors and leave the
program marked as SMP'able (i.e. tells the OS to let it run on multiple
processors), then it basically runs almost twice as fast (no suprise.)
If we leave the program marked as MP and turn off one processor, then
the first thread is about 1/3 faster, the second thread is the same as
the uniprocessor, the third thread is about 1/3 faster, and so on in an
interleaved manner. What would be the issue there? Seems kind of wierd.
(Nothing else going on and no interaction from us during the test.)

The other wierdness is that we can mark the program as uniprocessor
(tells OS/2 to run all its threads on 1 CPU), and all of the threads
except the last one run the same times. But the last one will often
take twice as long. That also seems a little funkey.

All the threads were left at normal priority and, since they are all
totally CPU bound, one would assume that they would basically get total
round robin dispatching regardless of the CPU setup.

This is hardly a major issue right now but I just though it was worth
a little pondering and might spark some lively debate. We will all soon
be dealing with this issue and I bet many people will find that their
threading strategy does not work quite the same on MP machines. If you
followed the rules, then the programs should work ok, but that is not
the same as working as fast as they can.

As to why our system applications run really bad, I have no idea. We
will have to look into it. We are running the shrink wrapped version of
OS/2 SMP, since the Dell machine is evidently a fully Intel SMP
compliant machine.

___
 X KWQ/2 1.2b X If at first you don't succeed, put out another version (KWQ 1.2).

--- Maximus/2 2.02