Slides available here!

Speaker:

Prof. Andreas Herkersdorf, TU Munich, Germany

Title:

What happens on an MPSoC stays on an MPSoC - unfortunately!

Abstract:

On today's and future MPSoC, consisting of dozens of diverse and freely programmable processor cores, it is hard to transform raw compute performance into system performance. A main reason is that programming massively parallel heterogeneous platforms is not yet sufficiently well understood.


Software diagnosis is the analysis whether software running on a specific hardware platform meets functional and non-functional requirements, i.e. whether the software is free of certain functional bugs and has sufficient performance characteristics. Two main prerequisites are necessary for thorough software diagnosis: an in-depth understanding of 1) how the software should execute on the target architecture, and 2) comprehensive run-time observations into the MPSoC showing how the software actually is executed. In order to understand how the software should be executed on an MPSoC, in depth knowledge of the architecture beyond the ISA boundary is necessary. This knowledge is highly specific to the MPSoC micro-architecture and typically only available to the processor OEM. The run-time observation of the software execution requires access to the full execution state (all CPU register content, memory content, NoC buffer content, etc.) of the target platform. With the growing number of integrated IP cores the execution state grows exponentially, amounting to multiple tens of Gbit/s up to Tbit/s today. At the same time, the available off-chip bandwidth for observing execution state is predicted to grow at most linearly from the multiple Mbit/s to single Gbit/s today. Already today, the observability gap is multiple orders of magnitude.


Traditional approaches to software diagnosis on MPSoCs, especially tracing, are not able to address the two prerequisites: The software developer usually knows too little about the MPSoC micro-architecture and the run-time observation fails due to insufficient I/O bandwidth.


In this talk, we will raise the awareness for the diagnosis challenges described above and present our new, work-in-progress approach to software diagnosis on MPSoCs. The approach stands on two conceptual pillars. First, we encode expert knowledge of the MPSoC architecture into inefficiency patterns and apply machine learning techniques to identify those patterns during program execution with corresponding temporal and spatial information. This information guides the software developer towards potentially relevant performance bottlenecks. Second, we avoid the I/O bandwidth bottleneck by detecting these patterns on-chip and constrain the off-chip communication to the obtained diagnosis information.


Bio:

Andreas Herkersdorf is a professor in the Department of Electrical and Computer Engineering and adjunct to the Department of Informatics at Technische Universität München (TUM). He received the Dipl.-Ing. degree from TUM in 1987 and the Dr. degree from ETH Zurich, Switzerland, in 1991, both in electrical engineering. Between 1988 and 2003, he has been in technical and management positions with the IBM Research Laboratory in Rüschlikon, Switzerland. Since 2003, Dr. Herkersdorf is director of the Chair for Integrated Systems at TUM. He is a senior member of the IEEE, member of the DFG (German Research Foundation) Review Board and serves as editor for Springer and Elsevier journals for design automation and communications electronics. His research interests include application-specific multi-processor architectures, IP network processing, Network on Chip, system level SoC modeling and design space exploration methods, and self-adaptive fault-tolerant computing.



* If you wish to modify any information or update your photo, please contact the web chairmpsoc2014@imag.fr