Zero-Copy Data Transfer for an OpenCL API Remoting System

In a cloud computing environment based on virtual machines, an OpenCL API remoting system forwards OpenCL commands issued by guest applications from the guest to the host. The forwarded commands are executed on the host side, and results are sent back to the guest. In this way, guest applications can make use of GPU devices installed on the host. If these commands require the transfer of data between guest applications and GPU devices, traditional API remoting systems usually copy the data between the guest and the host. The problem with this scheme is that the copying operation introduces additional cost and reduces data transfer bandwidth, which in turn decreases the overall performance of such systems. This paper utilizes a technique, called zero-copy data transfer, which allows sharing of memory between the guest and the host. Since the technique does not require to perform the copying operation, it can increase bandwidth significantly. With this technique, buffers of guest applications are first pinned to physical memory on the guest side, then mapped into some memory regions on the host side, and finally directly accessed by host programs. Our experiments demonstrated that the technique achieved a significantly higher transfer bandwidth than conventional network transfer. When the technique was utilized in an OpenCL API remoting system, for data-intensive OpenCL applications, overall performance was comparable to that of a physical OpenCL implementation.