Efficient CPU-GPU work sharing for data-parallel JavaScript workloads

Modern web browsers are required to execute many complex, compute-intensive applications, mostly written in JavaScript. With widespread adoption of heterogeneous processors, recent JavaScript-based data-parallel programming models, such as River Trail and WebCL, support multiple types of processing elements including CPUs and GPUs. However, significant performance gains are still left on the table since the program kernel runs on only one compute device, typically selected at kernel invocation. This paper proposes a new framework for efficient work sharing between CPU and GPU for data-parallel JavaScript workloads. The work sharing scheduler partitions the input data into smaller chunks and dynamically dispatches them to both CPU and GPU for concurrent execution. For four data-parallel programs, our framework improves performance by up to 65% with a geometric mean speedup of 33% over GPU-only execution.