GPU-Assisted Buffer Management

Cloud computing has become an emerging virtualization-based computing paradigm for various applications such as scientific computing and databases. Buffer management is an important factor for the I/O performance of the virtualized platform. In this study, we propose to leverage the memory and the computation power of the graphics processors (GPUs) to improve the effectiveness of buffer management. GPUs have recently been modeled as manycore processors for general-purpose computation. Designed as co-processors, they have an order of magnitude higher computation power than CPUs, and have a large amount of GPU memory, connected to the main memory with the PCI-e bus. In particular, we present two approaches of GPU-assisted buffer management, namely GRAM and DEDU. GRAM utilizes the GPU memory as additional buffer space and models the main memory and the GPU memory as a holistic buffer, whereas DEDU performs GPU-accelerated de-duplication to increase the effective amount of data pages that can fit into the extended buffer. We optimize both approaches according to the hardware feature of the GPU. We evaluate our algorithms on a workstation with an NVIDIA Tesla C1060 GPU using both synthetic and real world traces. Our experimental results show that the GPU-assisted buffer management reduces up to 68% of I/O cost for the traces generated from Xen.