PTask: Dataflow Programming for GPUs
Supporting future applications like gestural interfaces on low-power platforms such as cell phones and tablets forces developers to provide high compute bandwidth with low latency. To this end, PTask is a dataflow programming framework for GPUs that insulates the programmer from low-level details such as device-management, data transfer, and asynchrony. PTask is supported at the system call interface, so the OS can provide isolation and fairness guarantees for GPU computations.
Publications
- Christopher J. Rossbach, Jon Currey, Mark Silberstein, Baishakhi Ray, and Emmett Witchel, PTask: Operating System Abstractions To Manage GPUs as Compute Devices, Symposium on Operating Systems Principles (SOSP), 23 October 2011
- Christopher J. Rossbach, Jon Currey, and Emmett Witchel, Operating Systems must support GPU abstractions, in The 13th Workshop on Hot Topics in Operating Systems (HotOS), ACM, 8 May 2011
