Effective Dynamic Scheduling on Heterogeneous Multi/Manycore Desktop Platforms

Abstract

GPUs (Graphics Processing Units) have become one of the main co-processors that contributed to desktops towards high performance computing. Together with multicore CPUs and other co-processors, a powerful heterogeneous execution platform is built on a desktop for data intensive calculations. In our perspective, we see the modern desktop as a heterogeneous cluster that can deal with several applications' tasks at the same time. To improve application performance and explore such heterogeneity, a distribution of workload over the asymmetric PUs (Processing Units) plays an important role for the system. However, this problem faces challenges since the cost of a task at a PU is non-deterministic and can be influenced by several parameters not known a priori, like the problem size domain. We present a context-aware architecture that maximizes application performance on such platforms. This approach combines a model for a first scheduling based on an offline performance benchmark with a runtime model that keeps track of tasks' real performance. We carried a demonstration using a CPU-GPU platform for computing iterative SLEs (Systems of Linear Equations) solvers using the number of unknowns as the main parameter for assignment decision. We achieved a gain of 38.3% in comparison to the static assignment of all tasks to the GPU (which is done by current programming models, such as OpenCL and CUDA for Nvidia).

Extending the Scheduling Toolkit SEParAT to Support Hybrid Parallel Platforms

“…This implementation approach is supported by runtime systems like StarPU [6], Harmony [20] and Merge [21], and dynamic scheduling algorithms [22], [23]. In contrast to these approaches, SEParAT offers tool support for the static scheduling, which is especially beneficial for regular applications with a static task structure.…”

Efficient implementation of data flow graphs on multi-gpu clusters

“…Several task scheduling strategies for a multi-GPU platform based on a pool of tasks have been proposed [19,5,6,24]. However, the strategy of maintaining a pool of tasks that are randomly dispatched to idle processing units might perform poorly for applications with data dependencies and thus dependencies between tasks.…”

An Effective Dynamic Scheduling Runtime and Tuning System for Heterogeneous Multi and Many-Core Desktop Platforms

“…Consequently, we performed some improvements on the algorithm, leading to the Algorithm 2. The improvement is partially based on [20] and its main concept was shortly exposed on [21]. It performs a swap on pairs of assignment provided by the Algorithm 1 and verifies if such swap can promote a gain in the total performance of the tasks.…”