Choosing CPUs in an Open Market: System Performance Testing for the BaBar Online Farm

BABAR is a high-rate experiment to study CP violation in asymmetric e e collisions. The BABAR Online Farm is a pool of workstations responsible for the last layer of event selection, as well as for full reconstruction of selected events and for monitoring functions. A large number of machine architectures were evaluated for use in this Online Farm. We present an overview of the results of this evaluation, which include tests of low-level OS primitives, tests of memory architecture, and tests of application-specific CPU performance. Factors of general interest to others making hardware decisions are highlighted. Performance of current BABAR reconstruction (written in C++) is found to scale fairly well with SPECint95, but with some noticeable deviations. Even for machines with similar SPEC CPU ratings, large variations in memory system performance exist. No single operating system has an overall edge in the performance of its primitives. In particular, freeware operating systems perform no worse overall than the commercial offerings.