Procuring an HPC system - a.k.a. a supercomputer - is a complex and multifaceted task. Before sending out the Request For Proposals the procurer needs to quantify requirements along several dimensions and decide on an acceptable level of risk. Should the tendered system maximize benefit for existing users and use cases, or should possible future user communities and emerging HPC workloads be factored in? Do you prioritize throughput capability or minimizing time to solution for given workloads? How important are acquisition and running costs compared to other measures? Are you willing to invest in future technologies which would require significant refactoring of commonly used HPC simulation software? Which HPC software should be included in the benchmarking suite, and how should benchmark results be scored?
This session will start with a walk through of several aspects of an HPC procurement and will be followed by an open discussion where participants can share their own experiences. A goal of the discussion can be to arrive at a set of best practices in HPC procurements.
Q: How did your requirements gathering procedure work, how did you implement risk assesment to requirement process?
Q: Is the benchmark list exhaustive for the application benchmarks? Can you elaborate on why you chose such narrow scope of the benchmarks?
how much modification of benchmark codes should be allowed?
performance for individual proxy benchmarks and application benchmarks can point in different directions, but often becomes robust after averaging
to eliminate small/less competent vendors, can require certain minimum placement on top500 or alternatively have requirements on vendor stability or financial aspects