1. MPI ray tracer with data types. Modify the MPI-parallel master-worker style ray tracer so that it is not necessary for the master to copy the received tile data to the picture buffer explicitly. Hint: Use MPI derived data types.

  2. MPI-parallel dense MVM analysis. On the Emmy cluster, the Intel Trace Analyzer and Collector software is installed. It allows you to link your application with a special profiling library, run it, and view trace data graphically. Some documentation on how to use it can be found at https://hpc.fau.de/systems-services/systems-documentation-instructions/environment/#itac.  

    Use ITAC to analyze the MPI-parallel dense MVM code and confirm the hypotheses we made about its performance properties, specifically the claims about memory boundedness and communication overhead.

  3. Communication overhead and networks. How would the estimate for the MPI communication overhead in the dense MVM code change if Emmy had a bus network instead of a nonblocking fat-tree network? Assume that the basic point-to-point Ping-Pong performance characteristics (latency, asymptotic bandwidth) stay the same. 


Last modified: Thursday, 14 January 2021, 12:16 PM