High-Bandwidth Trace Collection for Multicomputer Performance Monitoring

Developing parallel applications requires sophisticated tools for tuning program performance. These tools require large quantities of data to be collected for analysis. A design for the collection and storage of trace data is presented in this paper. The design provides sufficient bandwidth and storage capacity for the collection of performance data from a 32--processor parallel computer using hybrid performance monitoring. A custom token--ring network is used to collect performance traces at a maximum bandwidth of 20 MBytes/sec. A 16 GByte disk array is utilized in the design for the storage of collected trace data. The resulting design provides a processor independent method for collecting performance data from moderately parallel computing systems. 1 Introduction Hybrid performance measurement combines the flexibility of software instrumentation with the unobtrusiveness of hardware monitoring to produce a viable technique for multicomputer performance monitoring [1][2]. Because o..