In the past decade, technologies such as non-volatile memory (NVM), high-bandwidth memory (HBM), compute express link (CXL), and near-memory computing (NMC) have dis- rupted established task and data placement heuristics. Con- ventional abstractions such as the memory hierarchy are no longer valid: for instance, using HBM as cache for DDR RAM can slow down workloads 8, 12. Task-based scheduling al- gorithms aim to resolve this by decomposing workloads into non-preemptible work units (tasks) and minimizing work- load execution time (makespan) by appropriate placement of tasks and data objects. However, their applicability to systems with novel, disruptive memory technologies such as HBM, CXL, or NMC depends on the underlying task and platform model. By examining the models used within hetero- geneous task-based scheduling algorithms published in the past two decades, we find that they have remained largely unchanged over the past 25 years, and are inappropriate for today’s level of heterogeneity in compute and memory components. Based upon this, and related work that offers at least partial improvements, we identify gaps in existing models, showcase examples to underline their relevance, and present ideas to remedy some of those. Our goal is not to give answers for all open questions, but rather to provide point- ers towards appropriate abstractions for future, operating system-centric, task-based scheduling research.
Friesel et al. (Thu,) studied this question.