Composing Multiple StarPU Applications over Heterogeneous Machines: A Supervised Approach

Abstract

Enabling HPC applications to perform efficiently when invoking multiple parallel libraries simultaneously is a great challenge. Even if a single runtime system is used underneath, scheduling tasks or threads coming from different libraries over the same set of hardware resources introduces many issues, such as resource oversubscription, undesirable cache flushes or memory bus contention.This paper presents an extension of StarPU, a runtime system specifically designed for heterogeneous architectures, that allows multiple parallel codes to run concurrently with minimal interference. Such parallel codes run within scheduling contexts that provide confined execution environments which can be used to partition computing resources. Scheduling contexts can be dynamically resized to optimize the allocation of computing resources among concurrently running libraries. We introduce a hypervisor that automatically expands or shrinks contexts using feedback from the runtime system (e.g. resource utilization). We demonstrate the relevance of our approach using benchmarks invoking multiple high performance linear algebra kernels simultaneously on top of heterogeneous multicore machines. We show that our mechanism can dramatically improve the overall application run time (-34%), most notably by reducing the average cache miss ratio (-50%).

Resource aggregation for task-based Cholesky Factorization on top of modern architectures

“…This contribution suffered from the fact that it does not allow to dynamically change the number of resources assigned to a parallel kernel. Our contribution in this study is a generalization of a previous work [16], where we introduced the so-called scheduling contexts which aim at structuring the parallelism for complex applications. Although our runtime system is able to cope with several flavors of inner parallelism (OpenMP, Pthreads, StarPU) simultaneously, in this paper we focus on the use of OpenMP to implement internal task parallelism.…”

“…The approach we propose in this paper to tackle the granularity problem is based on resource aggregation: instead of dynamically splitting tasks, we rather aggregate resources to process coarse grain tasks in a parallel manner on the critical resource, the CPU. To deal with Direct Acyclic Graphs (DAGs) of parallel tasks, we have enhanced the StarPU runtime system (see [8,16]) to cope with parallel tasks, the implementation of which relies on another parallel runtime system (e.g. OpenMP).…”

Multi-device Controllers: A Library to Simplify Parallel Heterogeneous Programming

“…More general approaches propose complete integrated frameworks that exploit lower-level specific programming models. Some examples include OM-PICUDA [9], Cashemere [6], StarPU [7] or the skeleton programming framework based on it, SkePU [2]. PACXX [4] is a transformation system integrated into the LLVM compiler framework.…”

Resource-Management Study in HPC Runtime-Stacking Context

“…StarPU relies on an hypervisor to dynamically chose which implementation of a kernel will be more suitable for the target hardware resources. Moreover, it uses a dynamic resource allocation with scheduling contexts [6]. The use of a global hypervisor may mitigate scheduling problems, however tasks are still coded with classic runtimes, thus creating situations where runtime stacking issues may appear.…”