The International Conference for High Performance Computing, Networking, Storage and Analysis
Scalable Client-Server Visualization of Callpath Traces for Large-Scale Parallel Executions.
Student: Philip A. Taffet (Rice University)
Supervisor: John M. Mellor-Crummey (Rice University)
Abstract: Scalable performance analysis tools are important for analyzing large-scale parallel executions. Sampling, typically far more scalable than instrumentation, still produces quantities of data that are unmanageable by current post-mortem analysis tools when analyzing large-scale executions. Hpctraceserver, a new addition to the HPCToolkit performance tools, improves the scalability of hpctraceviewer, a GUI post-mortem analysis tool, by extending it with a client-server model. Hpctraceserver, an MPI application, computes projections of the traces in parallel to fit the client’s window size with sampling and streams the results to the client over a TCP connection. Using the server allows hpctraceviewer to render views from remote multi-gigabyte databases in seconds. Hpctraceserver enables interactive analysis of large-scale parallel executions and removes the need to transfer large quantities of data in order to analyze them.