Lightweight Threading and Communication Shims for Chapel v. 1.0 Beta
This software extends the Cray Chapel runtime to use directly Sandia's Qthreads lightweight threading library and Sandia's Portals4 lightweight communication reference library.
SEARCH · Engineering Papers
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
This software extends the Cray Chapel runtime to use directly Sandia's Qthreads lightweight threading library and Sandia's Portals4 lightweight communication reference library.
In this paper, we evaluate the recently available directives in OpenMP 4 to parallelize a computational kernel using both the traditional shared memory approach and the newer accelerator targeting capabilities. In addition, we explore various transformations that attempt to increase application performance portability, and examine the expressiveness and performance implications of using these approaches. For example, we want to understand if the target map directives in OpenMP 4 improve data locality when mapped to a shared memory system, as opposed to the traditional first touch policy approach in traditional OpenMP. To that end, we use recent Cray and Intel compilers to measure the performance variations of a simple application kernel when executed on the OLCF s Titan supercomputer with NVIDIA GPUs and the Beacon system with Intel Xeon Phi accelerators attached. To better understand these trade-offs, we compare our results from traditional OpenMP shared memory implementations to the newer accelerator programming model when it is used to target both the CPU and an attached heterogeneous device. We believe the results and lessons learned as presented in this paper will be useful to the larger user community by providing guidelines that can assist programmers in the development of performance portable code.
Today s HPC centers user environments can be very complex. Centers often contain multiple large complicated computational systems each with their own user environment. Changes to a system s environment can be very impactful; however, a center s user environment is, in one-way or another, frequently changing. Because of this, it is vital for centers to notify users of change. For users, untracked changes can be costly, resulting in unnecessary debug time as well as wasting valuable compute allocations and research time. Communicating frequent change to diverse user communities is a common and ongoing task for HPC centers. This paper will cover the OLCF s current processes and methods used to communicate change to users of the center s large Cray systems and supporting resources. The paper will share lessons learned and goals as well as practices, tools, and methods used to continually improve and reach members of the OLCF user community.
We explore the trade-offs of performing linear algebra using Apache Spark, compared to traditional C and MPI implementations on HPC platforms. Spark is designed for data analytics on cluster computing platforms with access to local disks and is optimized for data-parallel tasks. We examine three widely-used and important matrix factorizations: NMF (for physical plausibility), PCA (for its ubiquity) and CX (for data interpretability). We apply these methods to 1.6TB particle physics, 2.2TB and 16TB climate modeling and 1.1TB bioimaging data. The data matrices are tall-and-skinny which enable the algorithms to map conveniently into Spark’s data parallel model. We perform scaling experiments on up to 1600 Cray XC40 nodes, describe the sources of slowdowns, and provide tuning guidance to obtain high performance.
The XGC fusion gyrokinetic code combines state-of-the-art, portable computational and algorithmic technologies to enable complicated multiscale simulations of turbulence and transport dynamics in ITER edge plasma on the largest US open-science computer, the CRAY XK7 Titan, at its maximal heterogeneous capability, which have not been possible before due to a factor of over 10 shortage in the time-to-solution for less than 5 days of wall-clock time for one physics case. Frontier techniques such as nested OpenMP parallelism, adaptive parallel I/O, staging I/O and data reduction using dynamic and asynchronous applications interactions, dynamic repartitioning.
NVRAM-based Burst Buffers are an important part of the emerging HPC storage landscape. The National Energy Research Scientific Computing Center (NERSC) at Lawrence Berkeley National Laboratory recently installed one of the first Burst Buffer systems as part of its new Cori supercomputer, collaborating with Cray on the development of the DataWarp software. NERSC has a diverse user base comprised of over 6500 users in 700 different projects spanning a wide variety of scientific computing applications. The use-cases of the Burst Buffer at NERSC are therefore also considerable and diverse. We describe here performance measurements and lessons learned from the Burst Buffer Early User Program at NERSC, which selected a number of research projects to gain early access to the Burst Buffer and exercise its capability to enable new scientific advancements. To the best of our knowledge this is the first time a Burst Buffer has been stressed at scale by diverse, real user workloads and therefore these lessons will be of considerable benefit to shaping the developing use of Burst Buffers at HPC centers.
This paper will discuss the entire process of acquiring and deploying Hopper from the first vendor market surveys to providing 3.8 million hours of production cycles per day for NERSC users. Installing the latest system at NERSC has been both a logistical and technical adventure. Balancing compute requirements with power, cooling, and space limitations drove the initial choice and configuration of the XE6, and a number of first-of- a-kind features implemented in collaboration with Cray have resulted in a high performance, usable, and reliable system.
Magnetic fusion is a long-term solution for producing electrical power for the world, and the large thermonuclear international device (ITER) being constructed will produce net energy and a path to fusion energy provided the computer modeling is accurate. To effectively address the requirements of the high-end fusion simulation community, application developers, algorithm designers, and hardware architects must have reliable simulation data gathered at scale for scientifically valid configurations. This paper presents detailed benchmarking results for a set of magnetic fusion applications with a wide variety of underlying mathematical models including both particle-in-cell and Eulerian codes using both implicit and explicit numerical solvers. Our evaluation on a petascale Cray XE6 platform focuses on profiling these simulations at scale identifying critical performance characteristics, including scalability, memory/network bandwidth limitations, and communication overhead. Overall results are a key in improving fusion code design, and are a critical first step towards exascale hardware-software co-design — a process that tightly couples applications, algorithms, imple- mentation, and computer architecture.
The XGC fusion gyrokinetic code combines state-of-the-art, portable computational and algorithmic technologies to enable complicated multiscale simulations of turbulence and transport dynamics in ITER edge plasma on the largest US open-science computer, the CRAY XK7 Titan, at its maximal heterogeneous capability, which have not been possible before due to a factor of over 10 shortage in the time-to-solution for less than 5 days of wall-clock time for one physics case. Frontier techniques such as nested OpenMP parallelism, adaptive parallel I/O, staging I/O and data reduction using dynamic and asynchronous applications interactions, dynamic repartitioning for balancing computational work in pushing particles and in grid related work, scalable and accurate discretization algorithms for non-linear Coulomb collisions, and communication-avoiding subcycling technology for pushing particles on both CPUs and GPUs are also utilized to dramatically improve the scalability and time-to-solution, hence enabling the difficult kinetic ITER edge simulation on a present-day leadership class computer.
We design and implement an efficient parallel approximation algorithm for the problem of maximum weight perfect matching in bipartite graphs, i.e. the problem of finding a set of non-adjacent edges that covers all vertices and has maximum weight. This problem differs from the maximum weight matching problem, for which scalable approximation algorithms are known. It is primarily motivated by finding good pivots in scalable sparse direct solvers before factorization where sequential implementations of maximum weight perfect matching algorithms, such as those available in MC64, are widely used due to the lack of scalable alternatives. To overcome this limitation, we propose a fully parallel distributed memory algorithm that first generates a perfect matching and then searches for weightaugmenting cycles of length four in parallel and iteratively augments the matching with a vertex disjoint set of such cycles. For most practical problems the weights of the perfect matchings generated by our algorithm are very close to the optimum. An efficient implementation of the algorithm scales up to 256 nodes (17,408 cores) on a Cray XC40 supercomputer and can solve instances that are too large to be handled by a single node using the sequential algorithm.
Now one of the fastest supercomputers in the United States, the Trinity Supercomputer Phase II project was completed during the summer of 2017, and the computer became fully operational during an unclassified “open science” run. The computer has now transitioned to classified mode. The Trinity supercomputer is designed to provide increased computational capability for the NNSA Nuclear Security Enterprise in support of increasing geometric and physics fidelities in nuclear weapons simulation codes, while maintaining expectations for total time to solution. The capabilities of Trinity are required for supporting the NNSA Stockpile Stewardship program’s certification and assessments to ensure that the nation’s nuclear stockpile is safe, secure and effective. With an expected operational lifetime through 2020, Trinity is an approximately 41.5 petaflop (41.5 million billion calculations per second, peak performance) supercomputer based on the Cray XC40 architecture. It consists of 19,420 compute nodes utilizing two processor architectures, Intel Xeon Haswell and Intel Xeon Phi Knights Landing. It is a water-cooled system with a power requirement of about 10 megawatts. The Trinity project is managed and operated by Los Alamos National Laboratory and Sandia National Laboratories under the Alliance for Computing at Extreme Scale (ACES) partnership. The system is located at the Nicholas Metropolis Center for Modeling and Simulation at Los Alamos and covers approximately 5,200 square feet of floor space.
Slated for delivery in 2021, Aurora (capable of a billion-billion calculations per second) is an exascale computing system built for AI that will usher in a new generation of computing power in the United States. Imagine a computer so powerful that it can map all the connections in the human brain, simulate every activity of a cancer cell or unlock the secrets of dark matter. That is just the beginning of the impacts Aurora will have our community, our region, our nation and the world. Argonne is partnering with the U.S. Department of Energy and industry collaborators Intel and Cray Inc. to build Aurora which will be housed at the Argonne Leadership Computing Facility, a DOE User Facility, in a new state-of-the-art building.
Automated fault prediction and diagnosis in HPC systems needs to be efficient for better system resilience. With increasing scalability required for exascale, accurate fault prediction aiding in quick remedy is hard. With changing supercomputer architectures, distilling fault data from the noisy raw logs requires substantial efforts. Predicting node failures in such voluminous system logs is challenging. To this end, we investigate an interesting way to pin-point node failures in such supercomputing systems. Our study on Cray system data with automated machine learning tools suggests that specific patterns of event messages on node unavailability can be indicator to node failures. This data extraction coupled with system and job data correlation helps in devising a methodology to predict node failures and their location over a specific time frame. This work aims to enable broader applicability for a generic fault prediction framework.
At the 2007 Cray User Group meeting, the Oak Ridge Leadership Computing Facility (OLCF) introduced the OLCF Test Harness (OTH), a framework[1] used for acceptance testing of the Jaguar supercomputer[2]. Since then, the OTH framework has evolved to version 2.0 which adds new features and streamlines usability. The OTH is the key piece of software used to orchestrate acceptance testing for all OLCF computational resources before they are deployed for production use, including our leadership class high performance computing (HPC) systems. The OTH framework is written in Python and is publicly available[3].In this paper, we first describe the requirements, design, and structure of the OTH. Then, we present specific improvements developed to support acceptance testing of the OLCF’s Summit system[4]. We will also showcase new OTH features that have been added to streamline the acceptance test process as well as the motivation behind those changes. As part of this work, we also evaluated different workflow tools in order to determine whether these tools could complement the OTH in two key areas: automation and reporting. The advantages and disadvantages identified with each tool will be discussed. Lastly, we summarize the challenges and lessons learned collected from using the OTH for the acceptance of the last three flagship systems at the OLCF. These may be useful for other HPC centers developing their own testing frameworks or those interested in using the OTH.
Heterogeneous systems are becoming increasingly prevalent. In order to exploit the rich compute resources of such systems, robust programming models are needed for application developers to seamlessly migrate legacy code from today’s systems to tomorrow’s. Over the past decade and more, directives have been established as one of the promising paths to tackle programmatic challenges on emerging systems. This work focuses on applying and demonstrating OpenMP offloading directives on five proxy applications. We observe that the performance varies widely from one compiler to the other; a crucial aspect of our work is reporting best practices to application developers who use OpenMP offloading compilers. While some issues can be worked around by the developer, there are other issues that must be reported to the compiler vendors. By restructuring OpenMP offloading directives, we gain an 18x speedup for the su3 proxy application on NERSC’s Cori system when using the Clang compiler, and a 15.7x speedup by switching max reductions to add reductions in the laplace mini-app when using the Cray-llvm compiler on Cori.
NERSC's newest system, Perlmutter, features a 35 PB all-flash Lustre file system built on HPE Cray ClusterStor E1000. We present its architecture, early performance figures, and performance considerations unique to this architecture. We demonstrate the performance of E1000 OSSes through low-level Lustre tests that achieve over 90% of the theoretical bandwidth of the SSDs at the OST and LNet levels. We also show end-to-end performance for both traditional dimensions of I/O performance (peak bulk-synchronous bandwidth) and non-optimal workloads endemic to production computing (small, incoherent I/Os at random offsets) and compare them to NERSC's previous system, Cori, to illustrate that Perlmutter achieves the performance of a burst buffer and the resilience of a scratch file system. Finally, we discuss performance considerations unique to all-flash Lustre and present ways in which users and HPC facilities can adjust their I/O patterns and operations to make optimal use of such architectures.
The Data Transfer Kit (DTK) is a software library designed to provide parallel data transfer services for arbitrary physics components based on the concept of geometric rendezvous. The rendezvous algorithm provides a means to geometrically correlate two geometric domains that may be arbitrarily decomposed in a parallel simulation. By repartitioning both domains such that they have the same geometric domain on each parallel process, efficient and load balanced search operations and data transfer can be performed at a desirable algorithmic time complexity with low communication overhead relative to other types of mapping algorithms. With the increased development efforts in multiphysics simulation and other multiple mesh and geometry problems, generating parallel topology maps for transferring fields and other data between geometric domains is a common operation. The algorithms used to generate parallel topology maps based on the concept of geometric rendezvous as implemented in DTK are described with an example using a conjugate heat transfer calculation and thermal coupling with a neutronics code. In addition, we provide the results of initial scaling studies performed on the Jaguar Cray XK6 system at Oak Ridge National Laboratory for a worse-case-scenario problem in terms of algorithmic complexity that shows good scaling on 0(1 x 104) cores for topology map generation and excellent scaling on 0(1 x 105) cores for the data transfer operation with meshes of O(1 x 109) elements. (authors)
This paper describes our experience with developing a parallel weighted- least-square (WLS) state estimation (SE) program for shared-memory parallel computers. Since the key computational kernel of the WLS algorithm based on the Newton-Raphson approach is a solver of sparse linear equations, a significant part of our effort was focused on selecting, implementing and evaluating this algorithm. An optimized shared memory version of the conjugate gradient (CG) algorithm was found to be competitive to state-of-the-art implementation of LU solvers for the SE problem on the SGI Altix, a shared memory architecture. We also ported the full SE algorithm including CG to the Cray MTA-2 shared memory multithreaded architecture and investigated its performance.