Network congestion is one of the biggest problems facing HPC systems today, affecting system throughput, performance, user experience, and reproducibility. Congestion manifests as run-to-run variability due to contention for shared resources (e.g., filesystems) or routes between compute endpoints. Despite its significance, current network benchmarks fail to proxy the real-world network utilization seen on congested systems. We propose a new open-source benchmark suite called the Global Performance and Congestion Network Tests (GPCNeT) to advance the state of the practice in this area. The guiding principles used in designing GPCNeT are described and the methodology employed to maximize its utility is presented. The capabilities of GPCNeT are evaluated by analyzing results from several world's largest HPC systems, including an evaluation of congestion management on a next-generation network. The results show that systems of all technologies and scales are susceptible to congestion and this work motivates the need for congestion control in next-generation networks.

Network congestion is one of the biggest problems facing HPC systems today, affecting system throughput, performance, user experience, and reproducibility. Congestion manifests as run-to-run variability due to contention for shared resources (e.g., filesystems) or routes between compute endpoints. Despite its significance, current network benchmarks fail to proxy the real-world network utilization seen on congested systems. We propose a new open-source benchmark suite called the Global Performance and Congestion Network Tests (GPCNeT) to advance the state of the practice in this area. The guiding principles used in designing GPCNeT are described and the methodology employed to maximize its utility is presented. The capabilities of GPCNeT are evaluated by analyzing results from several world's largest HPC systems, including an evaluation of congestion management on a next-generation network. The results show that systems of all technologies and scales are susceptible to congestion and this work motivates the need for congestion control in next-generation networks.

SummaryHermitian (symmetric) eigenvalue solvers are the core constituents of electronic structure, quantum‐chemistry, and other HPC applications such as Quantum ESPRESSO, VASP, CP2K, and NWChem to name a few. Our understanding of the performance of symmetric eigenvalue algorithms on various hardware is clearly important to the quantum chemistry or condensed matter physics community but in fact goes beyond that community. For instance, big data analytics is increasingly utilizing eigenvalues solvers, in the study of randomized singular value decomposition (SVD) or principal component analysis (PCA). Noise, vibration, and harshness (NVH) is another field where fast and efficient eigenvalue solvers are required. Most eigenvalue solver packages feature numerous different parameters which can be tuned for performance, eg, the number of nodes, number of total ranks, the decomposition of the matrix, etc. In this paper, we investigate the performance of different packages as well as the influence of these knobs on the solver performance.

We present an algorithm and implementation for the parallel computation of exact exchange in Quantum ESPRESSO (QE) that exhibits greatly improved strong scaling. QE is an open-source software package for electronic structure calculations using plane wave density functional theory, and supports the use of local, semi-local, and hybrid DFT functionals. Wider application of hybrid functionals is desirable for the improved simulation of electronic band energy alignments and thermodynamic properties, but the computational complexity of evaluating the exact exchange potential limits the practical application of hybrid functionals to large systems and requires efficient implementations. We demonstrate that existing implementations of hybrid DFT that utilize a single data structure for both the local and exact exchange regions of the code are significantly limited in the degree of parallelization achievable. We present a band-pair parallelization approach, in which the calculation of exact exchange is parallelized and evaluated independently from the parallelization of the remainder of the calculation, with the wavefunction data being efficiently transformed on-the-fly into a form that is optimal for each part of the calculation. For a 64 water molecule supercell, our new algorithm reduces the overall time to solution by nearly an order of magnitude.

NERSC has partnered with 20 representative application teams to evaluate performance on the Xeon-Phi Knights Landing architecture and develop an application-optimization strategy for the greater NERSC workload on the recently installed Cori system. In this article, we present early case studies and summarized results from a subset of the 20 applications highlighting the impact of important architecture differences between the Xeon-Phi and traditional Xeon processors. We summarize the status of the applications and describe the greater optimization strategy that has formed.

We develop a method for improving the parallel scalability of computations that involve asynchronous task execution. We apply this method to the recently developed parallel selected inversion algorithm [Jacquelin, Lin and Yang 2014], named PSelInv, on massively parallel distributed memory machines. In the PSelInv method, we compute selected elements of the inverse of a sparse matrix A that can be decomposed as A = LU, where L is lower triangular and U is upper triangular. Computing these selected elements of A-1 requires restricted collective communications among a subset of processors within each column or row communication group created by a block cyclic distribution of L and U. We describe how this type of restricted collective communication can be implemented using asynchronous point-to-point MPI communications combined with a binary tree based data propagation scheme. Because multiple restricted collective communications may take place at the same time, we need to use a heuristic to prevent processors participating in multiple collective communications from receiving too many messages. This heuristic allows us to reduce communication load imbalance and improve the overall scalability of the selected inversion algorithm. For instance, when 6, 400 processors are used, we observe that the use of this heuristic leads to over 5x speedup for a number of test matrices. It also mitigates the performance variability introduced by an inhomogeneous network topology.

We profile and optimize calculations performed with the BerkeleyGW [2,3] code on the Xeon-Phi architecture. BerkeleyGW depends both on hand-tuned critical kernels as well as on BLAS and FFT libraries. We describe the optimization process and performance improvements achieved. We discuss a layered parallelization strategy to take advantage of vector, thread and node-level parallelism. We discuss locality changes (including the consequence of the lack of L3 cache) and effective use of the on-package high-bandwidth memory. We show preliminary results on Knights-Landing including a roofline study of code performance before and after a number of optimizations. We find that the GW method is particularly well-suited for many-core architectures due to the ability to exploit a large amount of parallelism over plane-wave components, band-pairs, and frequencies.

Initial optimization strategies and results on MFDn, a large-scale nuclear physics application code, running on a single KNL node are presented. This code consists of the construction of a very large sparse real symmetric matrix and computing a few lowest eigenvalues and eigenvectors of this matrix through iterative methods. Challenges addressed include effectively utilizing MCDRAM with representative input data for production runs on 5,000 KNL nodes that require over 80 GB of memory per node, using OpenMP 4 to parallelize functions in the construction phase of the sparse matrices, and vectorizing those functions in spite of while-loops, conditionals, and lookup tables with indirect indexing. Moreover, hybrid MPI/OpenMP is employed not only to maximize the total problem size that can be solved per node, but also to eventually minimize parallel scaling overhead through the best scaling combination of MPI ranks per node with OpenMP threads. We describe a vectorized version of a popcount operation to avoid serialization on intrinsic popcnt which only operates on scalar registers. Additionally we leverage SSE 4.2 string comparison instructions to determine nonzero matrix elements. By utilizing MCDRAM, we achieve excellent Sparse Matrix–Matrix multiplication performance; in particular, using blocks of 8 vectors lead to a speedup of 6.4 × on KNL and 2.9 × on Haswell compared to the performance of repeated SpMV’s. This optimization was essential in achieving a 1.6 × improvement on KNL over Haswell.

We develop a method for improving the parallel scalability of the recently developed parallel selected inversion algorithm [Jacquelin, Lin and Yang 2014], named PSelInv, on massively parallel distributed memory machines. In the PSelInv method, we compute selected elements of the inverse of a sparse matrix A that can be decomposed as A = LU, where L is lower triangular and U is upper triangular. Updating these selected elements of A-1 requires restricted collective communications among a subset of processors within each column or row communication group created by a block cyclic distribution of L and U. We describe how this type of restricted collective communication can be implemented by using asynchronous point-to-point MPI communication functions combined with a binary tree based data propagation scheme. Because multiple restricted collective communications may take place at the same time in the parallel selected inversion algorithm, we need to use a heuristic to prevent processors participating in multiple collective communications from receiving too many messages. This heuristic allows us to reduce communication load imbalance and improve the overall scalability of the selected inversion algorithm. For instance, when 6,400 processors are used, we observe over 5x speedup for test matrices. It also mitigates the performance variability introduced by an inhomogeneous network topology.

The Gemini interconnect on the Cray XE6 platform provides for lightweight remote direct memory access (RDMA) between nodes, which is useful for implementing partitioned global address space (PGAS) languages like UPC and Co-Array Fortran. In this paper, we perform a study of Gemini performance using a set of communication microbenchmarks and compare the performance of one-sided communication in PGAS languages with two-sided MPI. Our results demonstrate the performance benefits of the PGAS model on Gemini hardware, showing in what circumstances and by how much one-sided communication outperforms two-sided in terms of messaging rate, aggregate bandwidth, and computation and communication overlap capability. For example, for 8-byte and 2KB messages the one-sided messaging rate is 5 and 10 times greater respectively than the twosided one. The study also reveals important information about how to optimize one-sided Gemini communication.

We present novel parallel language constructs for the communication intensive part of a magnetic fusion simulation code. The focus of this work is the shift phase of charged particles of a tokamak simulation code in toroidal geometry. We introduce new hybrid PGAS/OpenMP implementations of highly optimized hybrid MPI/OpenMP based communication kernels. The hybrid PGAS implementations use an extension of standard hybrid programming techniques, enabling the distribution of high communication work loads of the underlying kernel among OpenMP threads. Building upon lightweight one-sided CAF (Fortran 2008) communication techniques, we also show the benefits of spreading out the communication over a longer period of time, resulting in a reduction of bandwidth requirements and a more sustained communication and computation overlap. Experiments on up to 130560 processors are conducted on the NERSC Hopper system, which is currently the largest HPC platform with hardware support for one-sided communication and show performance improvements of 52% at highest concurrency.

Application codes in a variety of areas are being updated for performance on the latest architectures. In this paper we examine an application, which comes from magnetic fusion for performance acceleration with a particular emphasis on methods that are applicable for many/multicore and future architectural designs. We take an important magnetic fusion particle code that already includes several levels of parallelism including hybrid MPI combined with OpenMP. We study how to include new advanced hybrid models, which extend the applicability of OpenMP tasks and exploit multi-threaded MPI support to overlap communication and computation. Experiments carried out on Cray XT4 and XT5 machines resulting in a speed-up of up to 35% of the investigated GTS particle shifter kernel show the benefits and applicability of this approach.

: Application codes in a variety of areas are being updated for performance on the latest architectures. We describe current bottlenecks and performance improvement areas for applications including plasma physics, chemistry related to carbon capture and sequestration, and material science. We include a variety of methods including advanced hybrid parallelization using multi-threaded MPI, GPU acceleration, libraries and auto-parallelization compilers.

In the proceedings of CUG2006 we published a paper evaluating the performance of the HPCC benchmark suite on the Cray XT3, jaguar, and newly-upgraded Cray X1E, phoenix supercomputers at the Center for Computation Sciences (CCS) at Oak Ridge National Laboratory (ORNL). Although phoenix has remained unchanged since this paper was published, jaguar has undergone significant changes. It has been upgraded from a single-core XT3 system to a combined dual-core XT3 and XT4 system. This paper revisits HPCC on jaguar and discusses the effect that the past year's system upgrades have had on overall performance.

With the recent agreement to build the multi-billion dollar international burning plasma experiment known as ITER, fusion simulations will be growing dramatically in both complexity and size. The Gyrokinetic Toroidal Code (GTC) is a 3D Particle-In-Cell code used for studying the impact of finescale plasma turbulence on energy and particle confinement in the core of tokamak fusion reactors. To tackle global ITER-size simulations with full kinetic ion and electron physics, GTC will require new algorithms and supercomputers will have to grow in capabilities. We will review recent code modifications made to prepare for these new, exciting simulations and examine the performance of GTC on both the Cray XT3 and Cray XT4 systems using the latest Cray tools. Finally, we will look to the future and discuss plans for GTC and how that will effect its performance on future computers.

We describe a variety of optimizations of GTC, a 3D gyrokinetic particle-in-cell code, for the Cray X1E. Using performance data from multiple points in the optimization process, we compare the relative importance of dierent optimizations on the Cray X1E and their eects on Cray XT3 performance. We also describe the performance dierence between small pages and large pages on the XT3, a dierence that is particularly significant for GTC.

The last year has seen significant updates in the programming environment and operating systems on the Cray X1E and Cray XT3 as well as the much anticipated release of version 1.0 of HPCC Benchmark. This paper will provide an update and analysis of the HPCC Benchmark Results for Cray XT3 and X1E as well as a comparison against historical results.