Thread allocation is an important problem in distributed real-time and embedded (DRE) systems. A too liberal thread allocation policy may cause deadlock, a too conservative policy limits potential parallelism, thus wasting resources. However, achieving (global) optimal thread utilization, while avoiding deadlock, has been proven impractical in distributed systems: it requires too much communication between components.
In previous work we showed that efficient local thread allocation protocols are possible if the protocols are parameterized by global static data, in particular, an annotation of the global call graph of all tasks to be performed by the system. We proved that absence of cyclic dependencies in this annotation guarantees absence of deadlock.
In this paper we present an algorithm to compute optimal annotations, that is annotations that maximize parallelism while satisfying the condition of acyclicity. Moreover, we show that the condition of acyclicity is in fact tight and exhibits a rather surprising anomaly: if a cyclic dependency is present in the annotation of the call graph and a certain minimum number of threads is provided, deadlock is reachable. Thus, in the presence of cyclic dependencies, increasing the number of threads may introduce the possibility of deadlock in an originally deadlock free system.
In Proc. of the 20th IEEE Int'l Parallel and Distributed Processing Symposium (IPDPS'06)
Postscript, PDF. © 2006, Springer Verlag.