Mike Norman, right, director of the San Diego Supercomputer Center at UCSD, along with assistant director Allan Snavely look at Gordon, one of the most powerful computers in Southern California on Thursday, December 8, 2011.
Image Credit: © San Diego U-T via ZUMA Press Wire
Gordon Compute Node Board
Size: 10 x 28 in (25.4 x 71.2 cm)
Date: 2011
Manufacturer: APPRO International, Inc.
Artifact loan courtesy of the San Diego Supercomputing Center, University of San Diego
When a new computer system debuts in the upper 10% of the TOP500 list of the world’s fastest supercomputers, this success draws quick interest about the key design elements enabling this newcomer to reside among those fast machines. For the Gordon Supercomputer, which entered production at the San Diego Supercomputer Center in 2012 and was designed to be a “supercomputer system suitable for data-intensive applications” and not simply speedy calculating, its appearance on the TOP500 list generated more than the usual attention.
Gordon’s computational power came from 1,024 compute node boards like the one below, each having two Intel XEON E5-2670 CPUs running at 2.6 GHz. Their total computing capacity placed Gordon at the #48 spot on the November 2011 TOP500 list. By providing user programs with abundant CPU horsepower and a memory address space of up to 2TB, Gordon could readily support researchers working on the most demanding “big data” problems.
But where the system’s design truly excelled was in its ability to perform I/O, or input/output, operations. Gordon’s principal investigator Mike Norman demonstrated this at SC11 in Seattle where Gordon delivered a record-setting 36 million input/output operations per second owing to its pathbreaking design feature: 300TB of Intel SSD 710 series flash drives. According to Gordon’s co-principal investigator, Allan Snavely, “Intel SSDs were among the best” and easily helped to fulfill Gordon’s 2009 NSF award mandate “to bridge the widening latency gap between main memory and rotating disk storage in modern computing systems.”
Predictably, many researchers were eager to take advantage of Gordon’s promising performance realized through its design’s balanced combination of a flexible compute node architecture and its voluminous flash-based SSD storage.
For example, two users studying forest growth dynamics found Gordon’s ability to provide a 300GB shared aggregated memory space helped to reduce the final run time of their application from over a week to under eight hours.
Gordon’s low-latency SSD drives proved appealing to many users as either fast temporary storage spaces or as persistent storage for large databases. For instance, several computational chemists looking for new catalysts found that only Gordon offered the large scratch storage able to make their search for larger molecules by the CCSD(T) method practical. When it came to databases, calculations based on 3D molecule structures archived in the Protein Data Bank (PDB) saw a 4X speedup when the database was accessed from Gordon’s SSD versus a hard disk.
From 2012 to 2017 Gordon delivered 450 million core hours to these and over 9,000 other users who submitted jobs to support research from an extensive range of scientific disciplines. Some of these users were newcomers to supercomputing thanks to the outreach efforts made by the San Diego Supercomputer Center to promote Gordon and its data-intensive computing possibilities. While advanced hardware modules, such as this compute node board, helped to propel Gordon to a high TOP500 ranking, it was Mike Norman, Allan Snavely, and many other individuals who made it possible for these thousands of users to turn their data into knowledge.
