From: Ian King (iking_at_killthewabbit.org)
Date: Mon Jan 26 2004 - 17:21:43 PST
The authors create a taxonomy of thread management alternatives for
multiprocessors, assuming shared-memory implementations and providing
considerable low-level detail based on typical operating system structures, and
then analyze the performance implications of the resulting systems. This
discussion is supplemented by implementation of each alternative on an
appropriate platform, the Sequent Symmetry shared-memory multiprocessor.
Reading somewhat between the lines, it seemed the authors chose a user-mode
thread system, which was intriguing given the paper we read regarding scheduler
activations; between these two papers, we see the use of user-mode threads over
kernel-owned threads in multiprocessor environments, and significant performance
gains as a result.
The authors started from an optimized position - a user-mode threading
implementation - and analyzed how variations in the implementation could affect
performance; part of the argument was that by minimizing the overall overhead of
the threading mechanism, the performance implications of those variations would
be much more apparent. The authors seemed to find that the management of
locking for shared resources, and the resulting bus contention in various
shared-bus schemes, was the predominating factor in performance, with other
thread overhead being significant but secondary. The analysis of resulting
performance was quite rigorous in comparison with other papers we've read, and
the authors were quite thorough in addressing possible alternative
interpretations of the data.
I was particularly intrigued by the idle queue for processors; treating
processors as a resource is an idea I've considered in the context of the
exokernel. It is often implicitly assumed that the processor is the Lord and
Master of a computing system; while it is true that nothing happens without the
processor, from the perspective of tasks in a multitasking system that processor
is simply another resource to be shared. I think the idea of distinguishing
the services of a processor as a shareable resource from the perspective of the
scheduler adds considerable clarity to considerations of multiprocessor OS
implementation.
This archive was generated by hypermail 2.1.6 : Mon Jan 26 2004 - 17:26:19 PST