Sept. 15, 2021 — In August 2021, TACC hosted its 6th Texascale Days event: an opportunity for researchers to use the Frontera supercomputer at full (or half) scale to run computing jobs that cannot be performed anywhere else in the world.
Texascale Days primarily enables full-day production science runs; however, in May 2021 TACC added a benchmarking day as part of the event.
“This benchmarking opportunity helps researchers answer questions like: ‘Can I run the problem at this scale? Does it work? Is it worth it?'” said John Cazes, TACC’s director of HPC and the organizer of the event. “Researchers use this time to think in more detail about what they want to run at this scale for 24 hours.”
Six science teams participated in the latest benchmarking day. Working in two-hour blocks, the teams tested their codes at up to 8,192 nodes – roughly 450,000 cores — to determine how a code performs at scale, or how improvements to the physics or computing schemes impact its performance.
Simulating Quantum Devices
Brian La Cour, who directs the Applied Research Lab (ARL)’s UT Center for Quantum Research, participated in the most recent Texascale Days event. La Cour is working with Mark Selover (UT Physics), Brajesh Gupt (TACC), and Noah Davis (ARL) on large-scale quantum computer simulations as part of a Frontera Pathways project.
Quantum computers promise to revolutionize computing yet are still in the early stages of development. Classical simulation on high-performance computers provides a means to validate noisy, intermediate-scale quantum (NISQ) devices, but this is challenging for large problems.
“In 2016, for example, a 40-qubit simulation was performed on Stampede1 using the qHipster software package on 1,024 nodes,” La Cour explained. “For each additional qubit, the memory required for simulation doubles.”
La Cour and his team used 8,192 nodes on Frontera to perform a simulation of a 45-qubit random circuit similar to the one used in Google’s 2019 quantum supremacy experiment. The simulations were performed using the Quantum Exact Simulation Toolkit (QuEST) developed by their collaborators from the University of Oxford.
“With some custom modifications to the software developed by Mark Selover, we believe we can push this to 46 qubits in time for the next Texascale event,” La Cour said. “To our knowledge, that would be the largest random circuit simulation in the world.”
Trial and Error, at Scale
Michael Norman, a professor of Physics at the University of California, San Diego, used his two-hour test slot to see whether Enzo-E — a branch of the community-developed adaptive mesh refinement simulation code enabling multi-physics hydrodynamic astrophysical calculations — would scale up to 4,000 nodes with all of its physics and I/O capabilities engaged.
“We encountered some difficulties, but in the end were able to run the code without crashing, so a partial success,” Norman said. Despite the challenges, “this test time was just what we needed to prepare for a full physics production run later this fall.”
Andre Merzky, a senior research programmer at Rutgers University, experienced another useful trial and error situation using Frontera. He applied for a two-hour scaling slot to investigate a temporary slowdown his team had experienced during their last full-scale production run.
“We suspected the shared file system to be the cause of the issue,” he said. “We were successful in reproducing the problem, but made limited progress in understanding the underlying causes, so more work will be needed to resolve it.”
TACC experts provided him with system performance metrics for this run.
“Our numerical performance seems not to be as good as it could be,” Merzky wrote. “That in itself may not be much of a problem, but we’re happy to have those metrics to perform some cross-checking with expected and measured performance.”
More immediately successful were efforts by Hyungjun Lee and Feliciano Giustino of the Oden Institute at UT Austin to test the parallel performance of SternheimerGW, an open-source electronic structure software developed by the group used for calculating the excited-state properties of materials, such as Bi2Se3 (bismuth selenide), the prototypical topological insulator.
For Texascale Days, the team introduced OpenMP parallelization via multi-threading libraries. Also, “in order to minimize I/O impact on the file system, we implemented the low I/O mode which turns out to be essential in very large-scale runs,” Lee said.
With these simple modifications, the team demonstrated strong-scaling performance reaching 76 percent of the ideal speedup on up to 458,752 cores.
Extending Gravitational Wave Detection
The very first gravitational waves detected by LIGO in 2015 originated from two black holes spiraling around each other and merging together to form a single black hole. However, scientists believe some gravitational wave sources are not detectable by even the most advanced ground-based detectors. These include black hole binary mergers where the mass ratio of the system is in the range of 1:100 to 1:1000.
A careful comparison of a gravitational wave signal with a library of possible waveforms constructed from numerical simulations allows scientists to untangle information about a gravitational wave’s source.
“Numerical simulations of large mass ratio binaries are extremely computationally expensive and need scalable algorithms with spacetime adaptivity,” said Milinda Fernando, a post-doctoral fellow at the Oden Institute. “Existing codes for numerical relativity and relativistic magnetohydrodynamics do not scale well on modern heterogeneous clusters, which is a major impediment towards scientific progress in these areas.”
To overcome these challenges, the Dendro-GR simulation framework was developed by a computational team lead by Hari Sundar, Associate Professor at the School of Computing, University of Utah. During the recent Texascale Days, Fernando scaled up Dendro-GR to 4,096 Frontera nodes with excellent parallel efficiency. For the study, they used approximately 500,000 unknowns per core, with the largest problem containing 118 billion unknowns on the grid.
“These algorithmic enhancements allow it to be far more scalable than any existing such code,” said Omar Ghattas, chair in Computational Geosciences at UT Austin, chief scientist on Frontera, and a collaborator on the project. “Our interest moving forward is to augment the forward modeling capabilities of Dendro-GR with an inversion capability to provide higher fidelity gravitational wave detections.”
New Nuclei Studies
While most teams used their time to test code scaling, others pushed their computational studies further, enabling new and exciting discoveries.
Grigor Sargsyan, a graduate student at Louisiana State University in the group of Assistant Professor Kristina Launey, performed calculations during Texascale Days aimed at obtaining descriptions of several nuclei with astrophysical significance.
They calculated the wavefunctions of two isotopes — 7Li and 7Be — that play a vital role in the creation of heavier elements in the early stages of our universe. They also modeled the most common oxygen isotope, 16O — “the third most abundant in our solar system after hydrogen and helium, whose importance for the existence of life cannot be overstated,” Sargsyan said.
[The team previously simulated another alpha-conjugate nucleus, 8Be, on Frontera, with the results submitted to Physical Review Letters in August 2021 (preprint available at https://arxiv.org/abs/2107.10389).]
Descriptions of alpha-conjugate nuclei (nuclei with multiples of two protons and two neutrons) are challenging to derive from first principle approaches. However, it is possible to derive them with modern multi-petascale supercomputers like Frontera, according to Sargsyan — assuming you can access a large enough portion of the full system.
“With the current limit of 2,048 nodes for general production on Frontera, we are unable to run calculations for these isotopes in large enough model spaces,” said Sargsyan. “Texascale Days allows us to perform calculations in ultra-large model spaces required to describe this challenging process.”
From first-ever software runs, to scaling studies, to science at extreme scale, Texascale Days is providing parallel computing leaders and software developers the opportunity to push the limits of their codes.
“We’re preparing the community for future runs on Frontera and the exascale systems of tomorrow, while stress testing our system and giving us insights into how to improve performance,” Cazes concluded.
For full article and graphics, click here.
Source: Aaron Dubrow, Texas Advanced Computing Center