We propose a new operating system infrastructure to support direct access to network and storage I/O services from programmable computational accelerators and application processors, such as GPUs and DSPs. Together with improved programmability and system performance, this architecture eliminates the use of CPUs only for controlling I/O between peripherals and accelerators, thereby allowing significant energy savings.
Our work focus is on the software infrastructure and hardware support for:
1 .Direct access to storage and network services from DSPs
2. Direct control of network devices and network support for general-purpose GPUs We will investigate cross-layer software and hardware implications of supporting native I/O from accelerators.
We are developing a software prototype and will analyze its performance and energy efficiency on top of real SoCs with GPUs and DSPs using a set of representative media-intensive applications.