high performance computing on graphics processing units: hgpu.org

Many research works have recently experimented with GPU to accelerate packet processing in network applications. Most works have shown that GPU brings a significant performance boost when it is compared to the CPU-only approach, thanks to its highly-parallel computation capacity and large memory bandwidth. However, a recent work argues that for many applications, the key enabler for high performance is the inherent feature of GPU that automatically hides memory access latency rather than its parallel computation power. It also claims that CPU can outperform or achieve a similar performance as GPU if its code is re-arranged to run concurrently with memory access, employing optimization techniques such as group prefetching and software pipelining. In this paper, we revisit the claim of the work and see if it can be generalized to a large class of network applications. Our findings with eight popular algorithms widely used in network applications show that (a) there are many compute-bound algorithms that do benefit from the parallel computation capacity of GPU while CPU-based optimizations fail to help, and (b) the relative performance advantage of CPU over GPU in most applications is due to data transfer bottleneck in PCIe communication of discrete GPU rather than lack of capacity of GPU itself. To avoid the PCIe bottleneck, we suggest employing integrated GPU in recent APU platforms as a cost-effective packet processing accelerator. We address a number of practical issues in fully exploiting the capacity of APU and show that network applications based on APU achieve multi-10 Gbps performance for many compute/memory-intensive algorithms.