Avoiding unbounded priority inversion in barrier protocols using gang priority management

Abstract

Large real-time software systems such as real-time Java virtual machines often use barrier protocols, which work for a dynamically varying number of threads without using centralized locking. Such barrier protocols, however, still suffer from priority inversion similar to centralized locking. We introduce gang priority management as a generic solution for avoiding unbounded priority inversion in barrier protocols. Our approach is either kernel-assisted (for efficiency) or library-based (for portability) but involves cooperation from the protocol designer (for generality). We implemented gang priority management in the Linux kernel and rewrote the garbage collection safe-point barrier protocol in IBM's WebSphere Real Time Java Virtual Machine to exploit it. We run experiments on an 8-way SMP machine in a multi-user and multi-process environment, and show that by avoiding unbounded priority inversion, the maximum latency to reach a barrier point is reduced by a factor of 5.3 and the application jitter is reduced by a factor of 1.5.

Priority inversion with fungible resources

“…Assume that tasks are dependent upon events of other tasks, such as a high priority parent process that is waiting for a signal of termination from a low priority child process or a high priority Java thread that is waiting for a low priority thread to reach a barrier for garbage collection [15]. If the child process or the low priority Java thread is continuously preempted by independently executing, intermediate priority entities, it will indefinitely block the high priority entities.…”

“…In preemptive real-time systems, gang members are assigned affinities to different processors so that they execute in parallel whenever possible. The scheduler also chooses a gang priority to assign to all members of the same cooperating group [15]. Tasks that share resources, on the other hand, are assigned affinities to the same processor to limit competition over resources, with systems using priority scheduling and priority inheritance to control scheduling order at the processor.…”