Non-preemptive scheduling of real-time threads on multi-level-context architectures

Abstract

This paper addresses the problem of how to schedule periodic, real-time threads on a class of architectures referred to as multilevel-context (MLC) architectures. Examples of such architectures are real-time operating systems with support for user-or kernel-level threads, and multithreaded microprocessors endowed with on-chip contexts. A common feature of these architectures is that they provide support for the administration of threads within contexts at different levels of abstraction. Therefore, the cost for switching between threads will depend on the affinity of their corresponding contexts. The main contributions of this paper are to demonstrate (i) how the scheduling performance for off-line scheduling on MLC architectures can benefit from an integrated heuristic that is cognizant of both the time-criticality of a thread and the current context affinity; and (ii) how the predicted performance for on-line scheduling on MLC architectures can benefit from an off-line schedulability test that accounts for variations in the context affinity.