Chalk up another win for Sequoia and high-performance computing. The record-setting supercomputer is helping pave the way for future planetary-scale simulations.
Researchers from Lawrence Livermore National Laboratory (LLNL) and Renesselaer Polytechnic Institute (RPI) created a protocol called Time Warp that carried out 7.8 million MPI tasks on 1,966,080 cores of the Sequoia Blue Gene/Q supercomputer system.
Time Warp automatically exposes available parallelism in a model via its error detection and rollback recovery mechanism. The effect on performance is nothing short of remarkable: scaling from 32,768 cores to almost 2 million cores resulted in a 97x performance improvement. This long-sought-after linear performance gain is attributed to a sophisticated caching mechanism.
The team also used Sequoia to process 33 trillion events in 65 seconds, which comes out to over 504 billion events/second – that’s 41 times better than the previous record of 12.2 billion events per second, which was set in 2009.
The records were set using the ROSS (Rensselaer’s Optimistic Simulation System) simulation package in tandem with the Time Warp synchronization algorithm. Results were communicated as Warp Speeds, a long-range metric devised by the researchers to describe the progress of event-rates over many years. The current best time is represented as Warp Speed 2.7, while Warp Speed 10.0 is not expected for another 150 years.
The work has planetary-scale implications:
“The significance of this demonstration is that direct simulation of ‘planetary scale’ models is now, in principle at least, within reach,” said LLNL’s David Jefferson, who developed the Time Warp synchronization algorithm.
“Planetary scale” here refers to models capable of accommodating the largest systems, like the global population or the number of hosts on the Internet.
This research was carried out while Sequoia was operating as an unclassified ‘early science’ service. Last month, the system completed its transition to classified computing in support of the Stockpile Stewardship Program.
The research team, including Peter Barnes, Jr. and David Jefferson of LLNL, and Chris Carothers and Justin LaPre of Rensselaer, are publishing their results as part of the ACM SIGSIM Conference on Principles of Advanced Discrete Simulation (PADS).
Careful readers may be wondering how Sequoia, a machine that is listed as having 1,572,864 cores could run an application using almost 2 million cores. LLNL’s news release and the associated research paper characterize Sequoia as a 120-rack, 25-petaflop (peak) system with a minimum of 1,966,080 cores. However, the system’s spec page and the latest TOP500 list refer to Sequoia as a 96-rack, 20-petaflop (peak) system with 1,572,864 cores. Had Sequoia been upgraded?
I contacted LLNL’s Donald B. Johnston, who authored this latest announcement, to inquire about the source of the discrepancy. Don explained that while Sequoia was in its unclassified testing phase, it was connected to a smaller system creating a unified 120-rack system. The original terms of the Sequoia contract specified a 96-rack system for classified research, so when the system transitioned to classified status last month, the 120-rack system was separated into Sequoia (96 racks) and Voltan (24 racks). Voltan remains an open science community resource.