Path-Synchronous Performance Monitoring in HPC Interconnection Networks with Source-Code Attribution
Workshop: The 8th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems (PMBS17)
Authors: Adarsh Yoga (Rutgers University, Hewlett Packard Enterprise)
Abstract: Performance anomalies involving interconnection networks have largely remained a “black box” for developers relying on traditional CPU profilers. Network-side profilers collect aggregate statistics and lack source-code attribution. We have incorporated an effective protocol extension in the Gen-Z communication protocol for tagging network packets in an interconnection network; additionally, we have backed the protocol extension with hardware and software enhancements that allow tracking the flow of a network transaction through every hop in the interconnection network and associate it back to the application source code. The result is a first-of-its-kind hardware-assisted telemetry of disparate, autonomous interconnection networking components with application source code association that offers better developer insights. Our scheme works on a sampling basis to ensure low runtime overhead and generates modest volumes of data. Simulation of our methods in the open-source Structural Simulation Toolkit (SST/Macro) shows its effectiveness - deep insights into the underlying network details to the developer at minimal overheads.