Summit, now the fastest supercomputer in the world, is quickly making its mark in science – five of the six finalists just announced for the prestigious 2018 Gordon Bell Prize used Summit in their work. That’s impressive given that Summit only began full operation in early summer. Also noteworthy is Summit’s heterogeneous architecture which leverages IBM’s Power9 CPU, Nvidia V100 GPUs, and fast interconnect technology from Mellanox to accommodate traditional simulation workloads as well as mixed-precision workloads associated with AI and data analytics.
By now, Summit needs little introduction having topped the most recent Top500 list. Located at the Oak Ridge Leadership Computing Facility (OCLF), it cost an estimated $200 million to build as part of the DoE CORAL procurement program. It’s been heralded as the world’s most powerful supercomputer at 200 petaflops theoretical peak for high-performance computing workloads and 3.3 peak exaops for emerging AI workloads. (Sierra, the similarly architected but somewhat smaller, 125 petaflops theoretical peak machine based at Lawrence Livermore National Laboratory, was also used in some of the cited research.)
“These Gordon Bell finalists are an encouraging preview of the challenges users will be able to tackle on Summit when formal allocation programs begin in 2019,” said OLCF director of science Jack Wells. “Of particular note is the system’s ability to handle large volumes of data at scale, whether that be processing and analyzing experimental data or training artificial intelligence software to carry out specialized tasks.”
Nvidia and IBM are, understandably, ecstatic over Summit’s progress.
In a Nvidia blog, product manager Geetika Gupta wrote, “The revolutionary accelerators enable multi-precision computing that fuses the highly precise calculations to tackle the challenges of high performance computing with the efficient processing required for deep learning…[H]alf of the six projects included NVIDIA researchers who were heavily involved with the code development and performance tuning.”
Dave Turek, IBM Cognitive Systems VP, said “IBM designed Summit and Sierra to be data-centric, heterogeneous systems that maximized data flow for optimal application performance. The industry-leading IO features of IBM POWER9 processors allow for data to flow in and out of Summit’s GPUs to achieve the unprecedented level of performance demonstrated by these Gordon Bell finalists.”
They can, perhaps, be forgiven a little excess enthusiasm. These machines are difficult to design and build. Clearly, Summit’s early success is more evidence that heterogeneous architectures that leverage accelerators are likely to dominate high-end computing going forward.
Here is a lightly edited excerpt from an OCLF article describing the finalists who used Summit in their research:
- “Genomics. An ORNL team led by computational systems biologist Dan Jacobson and OLCF computational scientist Wayne Joubert that developed a genomics algorithm capable of using mixed-precision arithmetic to attain exascale speeds. On Summit, the team’s Combinatorial Metrics application achieved a peak throughput of 2.36 exaops—or 2.36 billion billion calculations per second, the fastest science application ever reported. Jacobson’s work compares genetic variations within a population to uncover hidden networks of genes that contribute to complex traits. One condition Jacobson’s team is studying is opioid addiction, which has been linked to the deaths of more than 49,000 people in the United States in 2017.
- Earthquake Simulation. A team from the University of Tokyo led by associate professor Tsuyoshi Ichimura that applied artificial intelligence (AI) and mixed-precision arithmetic to accelerate the simulation of earthquake physics in urban environments. As cities continue to grow, preparedness and improved understanding of ground-shaking’s effects on buildings and urban infrastructure become increasingly important. On Summit, the Tokyo team expanded on its 2014 algorithm, which was also a Gordon Bell Finalist, to achieve a fourfold speedup and to couple the shaking of ground and urban structures during large earthquakes into the same simulation.
- Extreme Weather. A Lawrence Berkeley National Laboratory-led collaboration that trained a deep neural network to identify extreme weather patterns from high-resolution climate simulations.The team, led by Berkeley data scientist Prabhat, plans to use the AI software to predict how extreme weather is likely to change in the future. By tapping into the specialized tensor cores built into Summit’s NVIDIA GPUs at scale, the Berkeley team achieved a peak performance of 1.13 exaops, the fastest deep-learning algorithm yet reported. Though the team applied its work to climate science, many of its innovations can be adapted for other deep-learning applications.
- Materials Science. An ORNL team led by data scientist Robert Patton that scaled a deep-learning technique on Summit to produce intelligent software that can automatically identify materials’ atomic-level information from electron microscopy With advanced microscopes capable of producing hundreds of images per day, real-time feedback supplied by AI could give scientists the ability to fabricate materials at the atomic level. Scaled across 4,200 nodes, the team’s MENNDL algorithm achieved a speed of 152.5 petaflops with an estimated performance rate of 167 petaflops across the whole machine.
- Physics. A team from Lawrence Berkeley and Lawrence Livermore National Laboratories led by physicists André Walker-Loud and Pavlos Vranas that developed improved algorithms to help scientists predict the lifetime of neutrons and answer fundamental questions about the universe. The team built upon its previous work using lattice quantum chromodynamics—a numerical method for calculating the underlying physics of the subatomic particles that make up protons and neutrons. In addition to optimized GPU software, the team developed lightweight, application-agnostic management software capable of managing hundreds of thousands of tasks. Using GPU-accelerated systems Sierra at Lawrence Livermore and the OLCF’s Summit, the team was able to start 1,056 four-node jobs on 4,224 nodes in 5 minutes, achieving a machine-to-machine speedup of factors of 10 and 15, respectively, over the OLCF’s previous leadership-class system, Titan. The achievement supplies nuclear physicists with the necessary computational power to support the experimental search for new physics.”
We’d be remiss not to mention the sixth Gordon finalist; it’s from a group of researchers from China who developed a graph processing framework (ShenTu) adapted for use on HPC resources. Here is a description of that very impressive work (ShenTu: Processing Multi-Trillion Edge Graphs on Millions of Cores in Seconds) taken from the SC18 web site.
“DescriptionGraphs are an important abstraction used in many scientific fields. With the magnitude of graph-structured data constantly increasing, effective data analytics requires efficient and scalable graph processing systems. Although HPC systems have long been used for scientific computing, people have only recently started to assess their potential for graph processing, a workload with inherent load imbalance, lack of locality, and access irregularity. We propose ShenTu, the first general-purpose graph processing framework that can efficiently utilize an entire petascale system to process multi-trillion edge graphs in seconds. ShenTu embodies four key innovations: hardware specializing, supernode routing, on-chip sorting, and degree-aware messaging, which together enable its unprecedented performance and scalability. It can traverse an unprecedented 70-trillion-edge graph in seconds.”
But back to Summit. Wells shared with HPCwire some of the distinguishing advantages Summit provides generally and some of which were leveraged by the Gordon Bell finalists. He noted two of the teams “were highly targeting the system’s mixed precision capabilities. The finite-element application explored ways mixed precision can boost performance by minimizing communication.”
Wells singled out four areas where Summit stands out:
- “Because of the NVLink the users can use more system memory than they could on Titan. Connecting the Volta GPUs to the Power9 CPU using NVLink provides much higher bandwidth than possible with PCIe Gen4. NVLink provides enough bandwidth so that the three GPUs can saturate the Power9’s memory bandwidth. This enables app to use system memory in addition to the GPU’s HBM which is not practical on systems like Titan with PCIe-attached GPUs.
- “The burst buffers, a reliable, high-speed storage layer that sits between the machine’s computing and file systems, significantly benefitted some teams who used it as a read accelerator rather than a write accelerator. Machine learning applications are read-heavy, so duplicating and moving data to the node local scratch memory was much faster than having it in GPFS.”
- “With this generation of InfiniBand, Mellanox has vastly improved its Adaptive Routing which greatly reduces congestion and allows applications to scale better. Additionally, one of the teams extensively took advantage of Mellanox’s switch-based collective operations, which shaved significant time off synchronization operations that typically limit an application’s scalability.
- “The Volta’s high bandwidth memory is very important. Summit’s nodes have more HBM than any other comparable system, which will allow our users to solve Gordon-Bell sized problems.”
On the software side, Summit users benefit from OCLF’s past experience with accelerators. Wells, noted, “Summit, like Titan, is a GPU-based system. Previous efforts to port and optimize codes for Titan have been beneficial for helping get codes ready for Summit. However, the Summit node is more complex, for example having multiple GPUs per node and having new features such as burst buffers and the GPU tensor cores. Adapting to this new node architecture has required effort by the code teams.”
Another issue for the Gordon Bell users, said Wells, is “[They only] had access to our relatively small test-and-development file system, not the full production file system that is undergoing acceptance testing these days. So, they had to work around this limitation. Also, the system software was still undergoing testing and debugging, so these teams were helping us identify such shortcomings and fix them.”
Information about the Summit stack – which includes XL, GNU, LLVM, PGI and NVCC compilers, LMOD, Spectrum MPI, ESSL, CUDA, LSF and JSM – is available on the web. The operating system is Red Hat Linux.
Obviously, these are early day for Summit which is still under preparation for full acceptance testing said Wells: “Users currently do not have access to the system as we attempt to finish this task. The IBM system is planned to be made available to the research community through DOE’s user programs beginning with allocations made under the Innovative and Novel Computational Impact on Theory and Experiment (INCITE) user program that will start in January 2019.”
That doesn’t mean plans aren’t afoot. They are. “For the past three years, teams have been preparing their applications to run on Summit,” said Wells. “A selection of the principal investigators of these application readiness teams includes:
- Salman Habib of Argonne National Laboratory, whose team is modeling the large-scale structure and distribution of matter over the 13-billion-year lifespan of the universe.
- Dmytro Bykov of Oak Ridge National Laboratory, whose team aims to describe the electronic structure of large molecular systems using quantum chemistry techniques, with targeted applications that include pharmacology and nanotechnology.
- Abhishek Singharoy of Arizona State University, whose team is investigating the mechanics of a biological motor called ATP synthase in all-atom detail, a study which may aid the design of bioinspired clean energy technology.
- Gaute Hagen of Oak Ridge National Laboratory, whose team is calculating the forces within atomic nuclei to study phenomena such as neutrinoless double-beta decay, a hypothesized form of radioactive decay.
- Joe Oefelein of Georgia Tech, whose team is carrying out combustion simulations that closely match engine operating conditions to inform the design of fuel-efficient, low-emission engines.”
The Gordon Bell Prize winner will be announced at the SC2018 in Dallas in November; as you may know it’s awarded each year by the Association of Computing Machinery (ACM) to recognize outstanding achievement in high-performance computing. “The purpose of the award is to track the progress over time of parallel computing, with particular emphasis on rewarding innovation in applying high-performance computing to applications in science, engineering, and large-scale data analytics…Financial support of the $10,000 award is provided by Gordon Bell, a pioneer in high-performance and parallel computing,” says ACM.
Link to article on OCLF web site: https://www.olcf.ornl.gov/2018/09/17/uncharted-territory/
Link to Nvidia blog: https://blogs.nvidia.com/blog/2018/09/17/nvidia-volta-tensor-core-gpus-gordon-bell-finalists/