Consistency model for runtime objects in the Open Community Runtime

Abstract

Task-based runtime systems are seen as a way to deal with increasing heterogeneity and performance variability in modern high performance computing systems. The Open Community Runtime is an example of such runtime system, with the benefit of having an open specification that defines its interface and behavior. This allows others to study the design and create alternative implementations of the runtime system. However, it turns out that the consistency model that defines how the OCR programs are synchronized when executed in parallel is not sufficiently defined. In the following text, we complete the consistency model by filling in the missing pieces and extending the memory model given in the specification.

The OCR-Vx experience: lessons learned from designing and implementing a task-based runtime system

“…We were able to extend the OCR memory model [15] to cover all OCR objects. There is now a clear definition of the order in which the operations on these objects should happen, based on the happens-before relation.…”

“…We can also see if there is no such ordering between two operations, creating a race condition. While building the memory model, we also filled in some missing pieces in the way happens-before was defined [15] . All OCR-Vx implementations as well as XSOCR implementation adhere to this extended memory model.…”

“…We have targeted both shared-memory and distributed-memory parallel architectures, as well as some less common architectures like the Intel Xeon Phi (both in the KNC coprocessor [18] and KNL processor format [19]) and OpenCL accelerators. We have also made multiple contributions to the OCR specification and developed a detailed consistency model for OCR [15]. A lot of our later work focused on the way OCR can be efficiently executed on non-uniform memory architecture (NUMA) systems [20], since we have discovered that the aspects of the OCR design originally aimed at distributed memory systems open up some interesting possibilities also for shared-memory NUMA systems.…”