Performance Validation of Network-Intensive Workloads on a Full-System Simulator

Performance accuracy is a critical but often neglected aspect of architectural performance simulators. One approach to evaluating performance accuracy is to attempt to reproduce observed performance results from a real machine. In this paper, we attempt to model the performance of a Compaq Alpha XP1000 workstation using the M5 full-system simulator. There are two novel aspects to this work. First, we simulate complex TCP/IP networking workloads and use network bandwidth as our primary performance metric. Unlike conventional CPU-intensive applications, these workloads spend most of their time in the operating system kernel and include significant interactions with platform hardware such as the interrupt controller and network interface device. Second, we attempt to achieve performance accuracy without extremely precise modeling of the reference hardware. Instead, we use simple generic component models and tune them to achieve appropriate bandwidths and latencies. Overall, we were able to achieve reasonable accuracy even with our relatively imprecise model, matching the bandwidth of the real system within 15% in most cases. We also used profiling to break CPU time down into categories, and found that the simulation results correlated well with the real machine.