RPO: Runtime web server optimization under simultaneous multithreading

Multicore architectures commonly feature simultaneous multithreading (SMT), a hardware technology to improve the performance of multi-threaded applications, such as web servers. By studying a TPC-W testbed we observe that the performance of SMT for a multi-tier application strongly depends on the workload mix in execution in the system, thus prompting the need for smart management policies to decide when to enable or disable SMT. To tackle this problem, we propose the Runtime SMT Performance Optimizer (RPO), a module for the Apache web server that automates SMT activation and deactivation at runtime. Decisions rely on the estimated mix of requests in execution in the system and a classification of transactions based on historical data about the ability of each request to benefit from SMT. Experimental results indicate that RPO can gain up to 40% in request latency compared to the best static SMT configuration policy, at the expense of a small overhead of 0.6% utilization on average for each core.