Programming and Managing Resources on Accelerator‐Enabled Clusters
M. Mustafa Rafique, Ali Raza Butt, Dimitrios S. Nikolopoulos · 2017
This chapter explores system design alternatives for clusters with computational accelerators and capability-aware task scheduling strategies for large-scale data processing on accelerator-enabled clusters. It presents an implementation and adaptation of the MapReduce programming model for asymmetric clusters. The asymmetry of resources on accelerator-enabled clusters introduces imbalances in resource management and provisioning. Addressing those imbalances, while hiding the associated complexity from users, is key to achieving high performance and high productivity. While the potential of many-core accelerators to catalyze high-performance computing (HPC) systems and data centers is clear, attempting to integrate accelerators seamlessly in large-scale computing installations raises challenges, with respect to resource management and programmability. Accelerators provide much higher performance to cost ratio compared with conventional processors. Thus, a properly designed accelerator-based cluster has the potential to provide high performance at a fraction of the cost and operating budget of a traditional symmetric cluster.