Shadowfax: scaling in heterogeneous cluster systems via GPGPU assemblies

Systems with specialized processors such as those used for accel- erating computations (like NVIDIA's graphics processors or IBM's Cell) have proven their utility in terms of higher performance and lower power consumption. They have also been shown to outperform general purpose processors in case of graphics intensive or high performance applications and for enterprise applications like modern financial codes or web hosts that require scalable image processing. These facts are causing tremendous growth in accelerator-based platforms in the high performance domain with systems like Keeneland, supercomputers like Tianhe-1, RoadRunner and even in data center systems like Amazon's EC2.

The physical hardware in these systems, once purchased and assembled, is not reconfigurable and is expensive to modify or upgrade. This can eventually limit applications' performance and scalability unless they are rewritten to match specific versions of hardware and compositions of components, both for single nodes and for clusters of machines. To address this problem and to support increased flexibility in usage models for CUDA-based GPGPU applications, our research proposes GPGPU assemblies, where each assembly combines a desired number of CPUs and CUDA-supported GPGPUs to form a 'virtual execution platform' for an application. System-level software, then, creates and manages assemblies, including mapping them seamlessly to the actual cluster- and node- level hardware resources present in the system. Experimental evaluations of the initial implementation of GPGPU assemblies demonstrates their feasibility and advantages derived from their use.