Faculty contributions drove this year’s increase in computational capacity on Savio, the Berkeley Research Computing (BRC) Program’s shared High Performance Computing (HPC) cluster. Condo contributions, including those expected to be added by the end of the year, totaled 63 nodes and 1,368 cores: a capacity expansion of approximately 19%, valued at about $400,000. Contributing researchers included faculty from Chemistry, Integrative Biology, Molecular and Cell Biology, Physics, and the Goldman School of Public Policy (GSPP). In addition, the Statistical Computing Facility (SCF) and the Computational Genomics Resource Laboratory (CGRL) each contributed nodes to Savio in 2016.
Among these diverse contributions, physics Professor Jeffrey Neaton, who is also Director of the Molecular Foundry, a nanoscience research center at Lawrence Berkeley National Laboratory, augmented his prior contribution to Savio with 16 Savio2 nodes (384 cores), cementing his status as the largest single condo contributor. Specialized Big Memory nodes were contributed by Ron Cohen (Chemistry), Solomon Hsiang (GSPP), and Daniel Weisz (Physics); a BigMem contribution by Doris Bachtrog (Integrative Biology) is currently in procurement. CGRL contributed High Throughput Computing (HTC) nodes.
Gary Jung, who leads HPC Services term, explains that BRC’s institutionally-funded investments this year expanded the infrastructure in which both condo and institutional nodes are installed in order to meet increasing demand: “BRC purchased five more racks to hold condo nodes,” he said, “as well as more infiniband infrastructure to support additional nodes, and a metadata storage array to speed metadata transactions on the parallel filesystem.” The BRC Program also purchased eight institutionally-funded nodes during the summer to augment the BigMem pool.
"One of Professor Solomon Hsiang's research interests is climate data analysis," says Sergey Shevtchenko, IT Director at the Goldman School of Public Policy. "This work requires analyzing terabytes of data using custom Python code, and running it on your average, run-of-the-mill server just wasn't fast enough. He's been looking move his research into a parallelized, supercomputer environment, so taking advantage of the Savio cluster at Berkeley was the natural solution."
Asked to describe how researchers are utilizing the Computational Genomics Research Laboratory’s contribution of HTC nodes, CGRL Director Jason Huff explains: "Several algorithms we use in genomics work on only one or a small number of processors, and the memory needed to compute on modern genomic datasets is typically in the tens of gigabytes. So, the HTC nodes are perfect because of the faster processors with a reasonable amount of memory. It is also provides CGRL users some familiarity, as the HTC nodes have core-level job allocation like our original cluster, Vector, providing users the flexibility to run single processes quickly or many serial processes one after another. Finally, there are enough nodes that users still have access to many processors in the condo pool when needed for larger jobs."
BRC’s Condo Cluster Program page on the Research IT website explains how the program works, and the characteristics of nodes that can be contributed. Interested researchers can e-mail research-it@berkeley.edu to set up a discussion about becoming a condo contributor to Savio.