| In cluster environments, all hosts are connected to be a distributed system by the high-speed network; the network hardware usually provides very low latency and high bandwidth. But traditional communication protocols introduce the overheads of process context switch, memory copies and frequent interrupts in the critical data path, the above overheads make the communication software unable to sufficiently deliver the hardware performance to user processes.This thesis illustrates the implementation as well as the performance analysis and optimization of user-level communication library. The main idea of user-level communication is bypassing the OS kernel when passing messages, so the software processing that is formerly executed in the kernel mode is now implemented in the user space. By creating the address mapping properly, user process is allowed to access the registers on network interface directly, in the mean time, the network interface can access the data in the user space straightway.In the implementation of our communication library, we also reduce the frequency of interrupts during communication. To achieve this goal, the"pre-post"technique is adopted to avoid interrupts when receiving messages; the communication software and network interface cooperate to implement remote direct memory access operations in order to reduce interrupt frequency and memory copies in shared memory clusters.We implemented a user-level communication library and achieved excellent performance: that under the"PingPong"benchmark, the lowest one-way latency is about 5.59μs, the bandwidth use-ratio of I/O bus is above 88%.We optimized the preliminary communication performance, and the optimized latency is 0.7μs lower on the whole, the bandwidth use-ratio of I/O bus is now above 94%. This thesis shows that, in the shared memory cluster environment, user-level communication mechanism can effectively deliver the high-performance of networking hardware to upper applications. |
PDF Full Text Request