The Berkeley Research Computing (BRC) Program is significantly expanding the capacity of the Savio supercomputing cluster, and is adding new computing models to support additional types of workflows across a broad range of research domains. The BRC condo program has seen rapid adoption, with campus partners doubling the size of Savio in the first 18 months. The number of users on Savio has jumped as additional researchers take advantage of the Faculty Computing Allowance launched this past Spring.
Expanded Savio capacity
Savio is adding over 20 million core-hours per year of compute capacity to campus researchers using the main compute cluster. This new compute pool consists of 104 latest-generation Intel Haswell processor nodes that augment the Intel IvyBridge nodes installed during the first phase of the program. Each new node provides about 70% more compute capacity than nodes in the previous generation (for more details on node specifications and the supporting infrastructure, please see Savio’s System Overview).
In addition, Savio is adding two new compute pools to address emerging needs for research computing:
- The Graphics Processing Unit (GPU) compute pool supports an increasing number of domains that employ highly parallel compute models, such as molecular and cell biology, as well as research requiring image processing. The GPU pool features 15 Intel Haswell processor nodes each equipped with two Nvidia Tesla K80 GPU cards. Taken together, the pool provides 60 GPUs with a total of nearly 150,000 compute cores.
- The High Throughput Computing (HTC) pool will be available for testing in December 2015. These nodes are also Haswell based, but are optimized for serial or single-core jobs, with fewer cores that run at higher speed. The HTC pool will provide a quicker resource response for serial computations by using a job scheduler policy that allows scheduling by-the-core (rather than by-the-node as is common for traditional HPC).
Growing breadth and depth of research support
A growing number of campus faculty are partnering with BRC in the Savio Condo program. As of October 2015, faculty in 16 departments and labs across campus have contributed a total of 160 compute nodes to Savio, with a combined value of over $900,000. The BRC program provides condo partners with infrastructure (racks, power & cooling, high-speed low-latency network, and fast parallel storage) and system administration, free of charge. In exchange, BRC can offer unused condo capacity to the rest of the campus community, spreading the benefit of research grants and departmental funding, and making research computing available to a broader population.
One of the ways BRC is making research computation broadly available to researchers is through the Faculty Computing Allowance. Launched in May 2015, the program has seen an enthusiastic response; over 60 faculty PIs have already brought their research groups into the program. Savio now has over 300 active users (well ahead of the program targets for adoption). Researchers using Savio under the Faculty Computing Allowance have run over 1.5 million core-hours of research computing tasks since the launch.
BRC recently piloted a program allowing condo partners to “burst” beyond the compute pool they contributed. A low-priority queue makes excess capacity available to condo research groups. Sometimes called a scavenging queue, this helps us to maximize utilization of Savio while ensuring that Faculty Computing Allowance users are not impacted. The pilot ran with several early partners, and will soon be made available to all condo users.
We anticipate continued growth in the condo program and the Faculty Computing Allowance, making high performance computing available across the campus research community. In addition, the BRC program provides Cloud Computing Consulting for those researchers whose needs are better met by commercial cloud providers (like AWS and Azure) and/or by national resources like the NSF-funded XSEDE program. BRC is also currently exploring needs for Analytics Environments on Demand (a.k.a. virtual workstations) to address social science and humanities workflows. These BRC programs, along with the recently launched Research Data Management (RDM) program, support a wide variety of compute- and data-intensive workflows to advance research at Berkeley.