Over the past few years, the NSF has funded a number of HPC systems to further supply the open research community with computational resources to meet that community’s changing and expanding needs. A review of these systems at the PEARC21 conference (July 19-22) highlighted the extreme diversity of architecture and hardware being explored to address a similar set of issues: handling ever-larger data; integrating and expanding the use of AI in scientific investigation; connecting what would once have been stand-alone HPC resources with the cloud and other external resources; and easing entry for research domains that never before needed computing, let alone HPC.
Moderated by Shawn Brown, director of the Pittsburgh Supercomputing Center (PSC) and Sergiu Sanielevici, director of user support at PSC, the panel explored eight new resources coming online throughout 2021. The systems fell under two NSF classifications: Category 1 systems provide production capabilities to the scientific community and are currently allocated by the NSF’s XSEDE cyberinfrastructure project. Category 2 systems are experimental deployments intended to explore new architectures and hardware that may form the basis for the next generation of HPC.
The PEARC conference series provides a forum for discussing challenges, opportunities and solutions among the broad range of participants in the research computing community. This community-driven effort builds on past successes and aims to grow and be more inclusive by involving additional local, regional, national and international cyberinfrastructure and research computing partners spanning academia, government and industry. PEARC21, Evolution Across All Dimensions, was offered this year as a virtual event (July 19-22).
Bridges-2: A Platform for Rapidly Evolving and Data-Intensive Research
Allowing users to explore the experimental space as never before is a central goal of PSC’s Bridges-2 advanced research computing platform, according to Brown, the project’s PI. An NSF Category 1 system (grant OAC-1928147) that entered its production phase earlier this year, Bridges-2 is meant to leverage Big Data. It is a heterogeneous system with a rapid interconnect and optimization for artificial intelligence to automate and innovate in computational experimentation.
Bridges-2 combines some disparate elements to allow portions of jobs to move among computing environments that best suit them and make the overall environment easy for users without prior HPC experience to enter:
- 488 “regular memory” nodes each with two AMD Epyc “Rome” CPUs and 256 GB RAM
- 16 “large memory” nodes with 512 GB RAM
- Four “extreme memory” nodes each with 4 TB of shared memory
- 24 GPU nodes with eight Nvidia Tesla V100 GPUs, two Xeon Gold CPUs and 512 GB RAM
- A Mellanox ConnectX-6 HDR InfiniBand 200 Gb/s interconnect
- Three levels of data storage, with >200 TB usable flash memory, a 15 PB usable Lustre file system and a ~17.5 PB compressed HPE StoreEver tape library
- Interactivity, containers, databases, gateways and compatibility with an array of popular languages and frameworks
The bottom line, Brown said, is enabling researchers to “get their work done” with a minimum startup effort.
“It’s not about how many boxes we can put in a room,” he said. “It’s about how much science we can accomplish.”
Allocations on the XSEDE-allocated Bridges-2 can be obtained through portal.xsede.org.
Neocortex, Unlocking Interactive AI Development for Rapidly Evolving Research
An example of an NSF Category 2 system at PSC is Neocortex (grant OAC-2005597), a specialized system leveraging Cerebras’ wafer-scale engine (WSE) technology to speed machine-learning training.
“State-of-the-art AI models are requiring more and more compute,” said Neocortex PI Paola Buitrago, director of AI & Big Data at PSC. “After 2012, we see that the compute required … is doubling every 3.4 months. Also, the models … just grow and get bigger and bigger” in terms of parameters and data. AI is also becoming more mainstream in science, expanding beyond its natural-language-processing and computer-vision beginnings into new domains such as biology.
Neocortex’s architecture is designed to accelerate training, the most time-consuming part of machine learning:
- A primary compute system consisting of two Cerebras CS-1 WSE servers, connected to an HPE Superdome Flex server each via 12 100Gb/s ethernet links
- Federation with the Bridges-2 platform and its 15 PB Lustre filesystem via 16 EDR-100 links
- 205 TB of NVMe SSD storage
- 24 TB RAM
To date, the unique high-memory architecture combined with the CS-1 servers, as well as the CS-1’s provision of 18 GB on-chip memory to its 400,000 AI-optimized core, has met the team’s expectations Buitrago said. “The fact that we have all this memory one clock away from the cores and also have a very fast fabric … enables very fast model-parallel training … It allows very high optimization of the hardware even at small batches.”
Updates on the system and information about joining the early user program can be found at www.cmu.edu/psc/aibd/neocortex/.
Exploring a Diverse Cyberinfrastructure with Ookami
Another NSF Category 2 testbed system, Ookami (grant OAC 1942140), is leveraging an architectural innovation pioneered by the fastest-in-world Fugaku system at RIKEN, said PI Robert Harrison, director of the Institute for Advanced Computational Science at Stony Brook University. The first deployment of the Fujitsu A64FX Post-K processor outside Japan, Ookami is exploring the use of Arm processors that promise “the performance of GPUs with the programmability of CPUs.”
Ookami’s Arm-based processors offer “a different path to computing at the leadership scale,” Harrison said, offering advantages in speed of memory, easily accessed performance and a fundamentally different path to exascale. Ookami features:
- 176 nodes, each with an A64FX processor and 32 GB (HBM) RAM
- A system total of 5.6 TB RAM
- An 0.86-PB Lustre filesystem storage
- An InfiniBand HDR-100 interconnect
- Scalable vector extensions, which enable the portability, scalability and optimization of vector length agnostic programming
- Predicate-centric architecture, supporting complex nested conditions and loops
- Support for open-source and commercial tools
Harrison explained that the 32-GB RAM was the product of a 2017 analysis of all jobs within the XSEDE system. It revealed that memory size would support 86 percent of XSEDE jobs and 85 percent of XSEDE cycles.
“Our experience with Ookami so far has been absolutely stellar,” he added, with various benchmark computations improving on Skylake CPU performance by 2.5 to 5.8 fold and excellent portability of code into the system. “All of these stacks essentially compile and run code essentially out of the box … It’s essentially a standard Linux environment.”
More information on Ookami can be found at www.stonybrook.edu/ookami/.
X. Carol Song, PI of the Category 1 Anvil system (OAC 2005632) and leader of the Scientific Solutions Group at Purdue University, introduced the upcoming system and its role in the XSEDE-allocated ecosystem.
“As people know, XSEDE … resources are always over-requested,” she said. With the evolving application domains and newer computational paradigms, she added, it will be challenging “to have enough resources to support those applications. And supremely important is the training of the next generation of researchers and workforce … Our answer is Anvil.”
Anvil is designed to address these issues with:
- High performance, from 1,000 compute nodes with AMD third-generation Epyc processors, with a peak performance of 5.3 PF, and 1 billion CPU core hours to XSEDE each year
- GPU/large-memory capabilities, with 16 GPU nodes, each with four Nvidia A100 GPUs, and 32 nodes of 1 TB memory each
- A composable subsystem featuring eight large-memory and storage nodes, integrated pathways to Microsoft Azure and Rancher to allow scalable deployment of Kubernetes infrastructure
- Multi-tier storage (including object storage, project and archival space), with 3 PB of flash storage, a 10 PB parallel file system and Globus data transfer
- Interactivity and compatibility with popular tools such as Jupyter Notebook, NGC containers, other AI-accelerated simulation toolkits, high-performance Python and R
Anvil’s earl user program will begin in mid-August this year, with a production target date of October 1. More information can be found at www.rcac.purdue.edu/anvil.
Jetstream2 as Part of the NSF Ecosystem
The upcoming, Category 1 Jetstream2 system (OAC 2005506) will build on experience with the current Jetstream, said PI David Hancock, director for Advanced Cyberinfrastructure at Indiana University.
Jetstream succeeded in allowing API access and full control to users, supporting “indefinite workflows” that allowed instances to run continuously, and developing successful trial applications, Hancock said. Jetstream2 will, in addition, provide small-allocation users with an expedited application process, multi-year allocations and shared dataset storage.
“It’s an evolution in multiple ways from the Jetstream system,” he said. With Jetstream 2, “we’re trying to emphasize aspects that we’ve already been providing and enhance those with heterogeneous hardware … and try to address some of the gaps we’ve found with Jetstream.”
Jetstream2 capabilities will include:
- An enhanced IaaS model with improved orchestration support, elastic virtual clusters, federated JupyterHubs, and improved storage sharing
- A commitment to >99 percent uptime to better support science gateways and hybrid-cloud computation
- A revamped user interface with unified instance management and multi-instance launch
- More than 57,000 next-gen AMD EPYC processor cores
- Over 360 Nvidia A100 GPUs
- Greater than 17 PB storage in an NVMe/disk hybrid
- A 100-GbE Mellanox network
Jetstream2 will continue Jetstream’s distributed model, with Internet2-linked regional sites at the University of Hawaii, Arizona State University, the Texas Advanced Computing Center and Cornell University in addition to the primary site at IU.
NSF has extended Jetstream’s operations by funding it through November of this year, with an extension through March 2022 in process. Early operations for Jetstream2 are planned for December 2021, with production operations beginning in January.
The National Center for Supercomputing Applications team chose the name “Delta” to signify the important changes they see in the advanced research computing field, said the upcoming Category 1 system’s (OAC 2005572) PI, Bill Gropp, director of NCSA.
“The system is called Delta because we’re looking to help the community be part of … the big change in what computing looks like going forward,” he said. “One of the things we’re seeing is that we’re in an era of great architectural innovation, and that also does come with challenges for the applications community to address.”
The largest GPU resource in the NSF ecosystem when it launches, Delta will contain:
- 124 CPU compute nodes and 8 utility nodes featuring AMD Epyc 7763 64-core Milan processors
- 100 each of 64- and 32-bit quad-GPU nodes with Nvidia A100 and A40 GPUs, respectively
- Five 8-way Nvidia A100 GPU nodes and one 8-way AMD MI100 GPU node, all of which do double-duty as high-memory nodes
- 3 PB relaxed-POSIX SSD-based storage and 7 PB of Lustre-based storage
Delta will support:
- Traditional HPC, data-intensive, and AI/ML workloads, with multiple compilers and runtimes
- Third-party software installation via Spack or Easybuild
- A modules utility for modifying programming environments
- Containers, including Singularity and vendor-supplied containers from Nvidia and AMD
- Persistent user tasks, supported by the utility nodes, for databases, workflow orchestration, etc.
- Gateways, through a key partnership with the Science Gateway Community Institute and engagement with multiple gateway communities
The Delta team will start deployment of the system this coming Fall. More information can be found about Delta at http://www.ncsa.illinois.edu/enabling/delta.
Voyager: Exploring AI Processors in Science and Engineering
Amit Majumdar, PI of the Category 2 Voyager system (OAC 2005369) and director, Data Enabled Scientific Computing at the San Diego Supercomputer Center, introduced the upcoming platform’s use of a unique computing resource: Habana Labs’ AI training and inference accelerators. Combined with Intel’s Xeon Scalable CPUs in Supermicro servers, the system will explore high-performance, high-efficiency AI-focused research in many domains.
“Voyager is an innovative resource for exploring AI processors in science and engineering,” Majumdar said. “It has specialized training and inference processors, and optimized implementations of standard machine learning frameworks. The high speed interconnect between training nodes provides a unique capability to scale up training models – something our research community is keen to explore.” Voyager will offer:
- 42 training nodes, each with eight Habana Gaudi processors, Ice Lake host CPUs and 6 TB of node-local storage
- Two inference nodes with eight Habana Goya processors and 3 TB of local storage
- 36 Intel x86 CPU compute nodes
- An on-chip, 400-GbE Arista Gaudi network
- A 3-TB high-performance storage system, accessible via a 25-GbE connection
- A 288-TB home file system, also connected via 25 GbE
- Deep-learning frameworks such as TensorFlow and PyTorch
- Software tools and libraries specific for Voyager’s architecture to enable user to develop AI techniques
Voyager’s delivery is expected in September of this year. The team is currently benchmarking the Gaudi training-node performance using remote access to units at Habana Labs, testing the Goya inference cards at the San Diego Supercomputer Center, and evaluating cnvrg.io for AI workflow. More information can be obtained by querying Majumdar at [email protected].
Expanse: Computing without Boundaries
The Category 1 Expanse project (OAC 1928224), headed by PI Mike Norman, offers a 5-petaflops HPC and data resource to the research community, according to co-PI Shawn Strande, deputy director of SDSC. The XSEDE-allocated system, which entered production operations in December 2020, is a heterogeneous HPC resource designed to support a broad spectrum of computational science and engineering research carried out at modest scale, but with increasingly diverse system and application software.
Integration with the outside world is a major focus of the system.
“We don’t want these machines to live in isolation,” Strande said. “The community expects and the science is driving systems to be better connected to external resources, whether that’s the public cloud, instruments, data sources, or integration with other scheduling infrastructure like the Open Science Grid … This is what we mean by ‘computing without boundaries’.”
- An HPC resource
- 13 non-blocking scalable compute units
- 728 CPU nodes with AMD Epyc Rome cores
- 93,184 cores in the full system, with 7,168 core in an HDR-100 non-blocking fabric
- 52 GPU nodes with 208 Nvidia V100 GPUs
- Four large-memory nodes
- Data-centric architecture
- 12 PB perf. Storage at 140 GB/s and 200K IOPS
- Fast I/O, node-local NVMe storage
- High-performance R & E networking
- Starting in August 2021, 7 PB of Ceph object storage
- Innovative operations
- Composable systems
- High-throughput computing
- Science gateways
- Interactive computing
- Integration with public cloud
Notably, Expanse features direct liquid cooling to boost reliability and performance.
“These are dense and power-hungry machines,” Strande said. “Moving forward it will be very difficult to air cool systems of similar design.”