Path: utzoo!attcan!uunet!sugar!ficc!walker
From: walker@ficc.uu.net (Walker Mangum)
Newsgroups: comp.os.misc
Subject: Multi-CPU system task scheduling
Keywords: Need help...
Message-ID: <2174@ficc.uu.net>
Date: 9 Nov 88 16:21:42 GMT
Organization: Adytum, Inc.
Lines: 44


I'm working on a real-time multi-CPU system in which a process can
be scheduled by another process, a timer, or an event in the "most
attractive" CPU.  The actual mechanics of loading and scheduling
the new process work just fine.  The problem is the definition of
"most attractive" CPU.

The goal of "most attractive" scheduling is to get the process to
execute in the CPU that will allow the process to complete as
quickly as possible, and to keep resource utilization among the
CPUs somewhat even.  Once a process is assigned to a CPU, it
remains in that CPU until it completes.  Process scheduling within
CPUs is priority based.

Currently, the "most attractive" CPU is determined by sampling the
load average for all CPUs, and selecting the CPU that has the
lowest load average and has no processes holding for memory
allocation.

THe problem is that the load average is determined once per second,
and occasionally *many* processes are simultaneously scheduled.
This means that, during each second, the current "most attractive"
CPU does not change (unless it becomes low on available memory),
and *all* processes are scheduled in that CPU!  Needless to say,
the results are not as desired....

We're considering adding an arbitrary delta to the load average for
a CPU when a process is scheduled that CPU.  This delta could be
(although not easily) a unique value for each process, with the
proper value for each process determined by system test and
performance evaluation.  This would effectively scatter processes
scheduled simultaneously, avoiding the current overloading problem.

Is there a better way to handle this problem?  I would appreciate
*any* and *all* ideas/answers.

Please to uunet!ficc!walker.

----------------------------------------------------------------------
-- 
Walker Mangum                                  |  Adytum, Incorporated
phone: (713) 333-1509                          |  1100 NASA Road One  
UUCP:  uunet!ficc!walker  (walker@ficc.uu.net) |  Houston, TX  77058
Disclaimer: $#!+ HAPPENS