Parallel and high-performance computing

A server receives bursts of short independent jobs. Why might a fixed-size worker thread pool be preferable to creating a new thread for every job?