The internal structure of interactions in a hidden network can be inferred using a maximum likeli... more The internal structure of interactions in a hidden network can be inferred using a maximum likelihood estimate based on<br> a record of its external behavior, within the framework of the kinetic Ising model. Beyond its origins in statistical physics,<br> solutions to this problem can model the internal structure of a hidden neural network based on activity recordings from a<br> laboratory setting, or the training process of an arti cial neural network in the context of machine learning. The primary<br> obstacle to its practical application is that the amount of computational work required grows rapidly with the dimensions<br> of the represented network, but the vast majority of the operations can be independently evaluated in parallel. In this<br> paper, we investigate the performance characteristics of a proxy application that models this growth, with the purpose<br> of examining its suitability as a candidate application for future exascal...
Computational Fluid Dynamics encompasses a great variety of numerical approaches that approximate... more Computational Fluid Dynamics encompasses a great variety of numerical approaches that approximate solutions to the<br> Navier-Stokes equations, which generally describe the movements of viscous<br> uid substances. While the objectives of these<br> approaches are to capture related physical phenomena, the details of di erent methods lend them to particular classes of<br> problems, and scalable solutions are important to a large range of scienti c and engineering applications. In this paper,<br> we investigate the practical scalability of two proxy applications that are made to recreate the essential performance<br> characteristics of Lattice-Boltzmann Methods (LBM) and Smoothed Particle Hydrodyamics (SPH), using the former to<br> simulate the formation of vortices resulting from sustained, laminar ow, and the latter to simulate violent free surface ows without a mesh. The di ering scalability properties of these methods suggest di erent desig...
Proceedings of the 25th European MPI Users' Group Meeting, 2018
Fortran coarrays are an attractive alternative to MPI due to a familiar Fortran syntax, single si... more Fortran coarrays are an attractive alternative to MPI due to a familiar Fortran syntax, single sided communications and implementation in the compiler. Scaling of coarrays is compared in this work to MPI, using cellular automata (CA) 3D Ising magnetisation miniapps, built with the CASUP CA library, https://cgpack.sourceforge.io, developed by the authors. Ising energy and magnetisation were calculated with MPI_ALLREDUCE and Fortran 2018 co_sum collectives. The work was done on ARCHER (Cray XC30) up to the full machine capacity: 109,056 cores. Ping-pong latency and bandwidth results are very similar with MPI and with coarrays for message sizes from 1B to several MB. MPI halo exchange (HX) scaled better than coarray HX, which is surprising because both algorithms use pair-wise communications: MPI IRECV/ISEND/WAITALL vs Fortran sync images. Adding OpenMP to MPI or to coarrays resulted in worse L2 cache hit ratio, and lower performance in all cases, even though the NUMA effects were rule...
One of the main hurdles of a broad distribution of PGAS approaches is the prevalence of MPI, whic... more One of the main hurdles of a broad distribution of PGAS approaches is the prevalence of MPI, which as a de-facto standard appears in the code basis of many applications. To take advantage of the PGAS APIs like GASPI without a major change in the code basis, interoperability between MPI and PGAS approaches needs to be ensured. In this article, we address this challenge by providing our study and preliminary performance results regarding interoperating GASPI and MPI on the performance crucial parts of the Ludwig application.
Accelerators and, in particular, Graphics Processing Units (GPUs) have emerged as promising compu... more Accelerators and, in particular, Graphics Processing Units (GPUs) have emerged as promising computing technologies which may be suitable for the future Exascale systems. Here, we present performanc ...
In radiotherapy, dose calculation is one of the most important and computationally intensive stag... more In radiotherapy, dose calculation is one of the most important and computationally intensive stages needed in a cancer treatment planning. During current clinical cancer treatment, the dose is calculated before the first treatment session and it will continue until the end of the therapy. However, tumours may change during the treatment and it is needed a re-calculation of the dose. Online Adaptive Radiotherapy (ART) is a technique that promises to deliver an optimal dose according to the necessity of the patient anatomy. Unfortunately, this technique is not possible yet due to technical limitations. One of those limitations is the time needed to recalculate the new dose to be delivered. Finite Size Pencil Beam algorithm seems to provide the requirements to fit in this technique as it is a fast, accurate enough and very suitable to run on graphics processing units (GPUs). Here we have investigated the advantages and capacities of using GPUs for cancer treatment planning based on a p...
With the arrival of “big data” (microarrays, screens, next-generation sequencing) in the life sci... more With the arrival of “big data” (microarrays, screens, next-generation sequencing) in the life sciences, standard analyses of these data for regular users of R now run into severe issues of computation time or computer memory. Many projects (including parallelisation efforts of the R core) offer R packages and functions that allow programming of solutions for large-scale analysis problems. However, these usually require familiarity with HPC programming as well as sufficient and funded time to employ, which is feasible for one-off analysis problems but impractical for common analysis methods.
One of the main hurdles of a broad distribution of PGAS approaches is the prevalence of MPI, whic... more One of the main hurdles of a broad distribution of PGAS approaches is the prevalence of MPI, which as a de-facto standard appears in the code basis of many applications. To take advantage of the PGAS APIs like GASPI without a major change in the code basis, interoperability between MPI and PGAS approaches needs to be ensured. In this article, we address this challenge by providing our study and preliminary performance results regarding interoperating GASPI and MPI on the performance crucial parts of the Ludwig and iPIC3D applications. In addition, we draw a strategy for better coupling of both APIs.
Accelerators and, in particular, Graphics Processing Units (GPUs) have emerged as promising compu... more Accelerators and, in particular, Graphics Processing Units (GPUs) have emerged as promising computing technologies which may be suitable for the future Exascale systems. Here, we present performance results of NekBone, a benchmark of the Nek5000 code, implemented with optimized OpenACC directives and GPUDirect communications. Nek5000 is a computational fluid dynamics code based on the spectral element method used for the simulation of incompressible flow. Results of an optimized NekBone version lead to 78 Gflops performance on a single node. In addition, a performance result of 609 Tflops has been reached on 16, 384 GPUs of the Titan supercomputer at Oak Ridge National Laboratory.
Coarrays are a native Fortran means for SPMD parallel programming, implementing a single-sided co... more Coarrays are a native Fortran means for SPMD parallel programming, implementing a single-sided communications model. Coarray Fortran belongs to the PGAS class of parallel languages. Relatively few tools support profiling of coarray programs. Profiling and tracing of pure coarray programs and mixed coarray+MPI miniapps was done in this work on 2 platforms. TAU (Tuning and Analysis Utilities) toolkit 2.25.2 was used with the Intel compiler 16.0.2 and the Intel MPI libraries 4.1.0 on Red Hat Enterprise Linux 6 cluster. CrayPAT (Performance Analysis Tool) was used on Cray XC30 system with CCE (Cray Compiler Environment) 8.4.1. Four coarray programs, of progressively increasing complexity, are studied. TAU profiling results on 32 cores and CrayPAT profiling data on 7k cores show that our hybrid MPI+coarray codes are relatively well balanced. We show that the Intel compiler maps coarray remote operations and synchronisation routines onto MPI-2 RMA calls. For the Intel implementation we sh...
We have developed miniapps from MPI finite element library ParaFEM and Fortran 2008 coarray cellu... more We have developed miniapps from MPI finite element library ParaFEM and Fortran 2008 coarray cellular automata library CGPACK. The miniapps represent multiscale fracture models of polycrystalline solids. The software from which these miniapps have been derived will improve predictive modelling in the automotive, aerospace, power generation, defense and manufacturing sectors. The libraries and miniapps are distributed under BSD license, so these can be used by computer scientists and hardware vendors to test various tools including compilers and performance monitoring applications. CrayPAT tools have been used for sampling and tracing analysis of the miniapps. Two routines with all-to-all communication structures have been identified a primary candidates for optimisation. New routines have been written implementing the nearest neighbour algorithm and using coarray collectives. Scaling limit for miniapps has been increased by a factor of 3, from about 2k to over 7k cores. The miniapps ...
The internal structure of interactions in a hidden network can be inferred using a maximum likeli... more The internal structure of interactions in a hidden network can be inferred using a maximum likelihood estimate based on<br> a record of its external behavior, within the framework of the kinetic Ising model. Beyond its origins in statistical physics,<br> solutions to this problem can model the internal structure of a hidden neural network based on activity recordings from a<br> laboratory setting, or the training process of an arti cial neural network in the context of machine learning. The primary<br> obstacle to its practical application is that the amount of computational work required grows rapidly with the dimensions<br> of the represented network, but the vast majority of the operations can be independently evaluated in parallel. In this<br> paper, we investigate the performance characteristics of a proxy application that models this growth, with the purpose<br> of examining its suitability as a candidate application for future exascal...
Computational Fluid Dynamics encompasses a great variety of numerical approaches that approximate... more Computational Fluid Dynamics encompasses a great variety of numerical approaches that approximate solutions to the<br> Navier-Stokes equations, which generally describe the movements of viscous<br> uid substances. While the objectives of these<br> approaches are to capture related physical phenomena, the details of di erent methods lend them to particular classes of<br> problems, and scalable solutions are important to a large range of scienti c and engineering applications. In this paper,<br> we investigate the practical scalability of two proxy applications that are made to recreate the essential performance<br> characteristics of Lattice-Boltzmann Methods (LBM) and Smoothed Particle Hydrodyamics (SPH), using the former to<br> simulate the formation of vortices resulting from sustained, laminar ow, and the latter to simulate violent free surface ows without a mesh. The di ering scalability properties of these methods suggest di erent desig...
Proceedings of the 25th European MPI Users' Group Meeting, 2018
Fortran coarrays are an attractive alternative to MPI due to a familiar Fortran syntax, single si... more Fortran coarrays are an attractive alternative to MPI due to a familiar Fortran syntax, single sided communications and implementation in the compiler. Scaling of coarrays is compared in this work to MPI, using cellular automata (CA) 3D Ising magnetisation miniapps, built with the CASUP CA library, https://cgpack.sourceforge.io, developed by the authors. Ising energy and magnetisation were calculated with MPI_ALLREDUCE and Fortran 2018 co_sum collectives. The work was done on ARCHER (Cray XC30) up to the full machine capacity: 109,056 cores. Ping-pong latency and bandwidth results are very similar with MPI and with coarrays for message sizes from 1B to several MB. MPI halo exchange (HX) scaled better than coarray HX, which is surprising because both algorithms use pair-wise communications: MPI IRECV/ISEND/WAITALL vs Fortran sync images. Adding OpenMP to MPI or to coarrays resulted in worse L2 cache hit ratio, and lower performance in all cases, even though the NUMA effects were rule...
One of the main hurdles of a broad distribution of PGAS approaches is the prevalence of MPI, whic... more One of the main hurdles of a broad distribution of PGAS approaches is the prevalence of MPI, which as a de-facto standard appears in the code basis of many applications. To take advantage of the PGAS APIs like GASPI without a major change in the code basis, interoperability between MPI and PGAS approaches needs to be ensured. In this article, we address this challenge by providing our study and preliminary performance results regarding interoperating GASPI and MPI on the performance crucial parts of the Ludwig application.
Accelerators and, in particular, Graphics Processing Units (GPUs) have emerged as promising compu... more Accelerators and, in particular, Graphics Processing Units (GPUs) have emerged as promising computing technologies which may be suitable for the future Exascale systems. Here, we present performanc ...
In radiotherapy, dose calculation is one of the most important and computationally intensive stag... more In radiotherapy, dose calculation is one of the most important and computationally intensive stages needed in a cancer treatment planning. During current clinical cancer treatment, the dose is calculated before the first treatment session and it will continue until the end of the therapy. However, tumours may change during the treatment and it is needed a re-calculation of the dose. Online Adaptive Radiotherapy (ART) is a technique that promises to deliver an optimal dose according to the necessity of the patient anatomy. Unfortunately, this technique is not possible yet due to technical limitations. One of those limitations is the time needed to recalculate the new dose to be delivered. Finite Size Pencil Beam algorithm seems to provide the requirements to fit in this technique as it is a fast, accurate enough and very suitable to run on graphics processing units (GPUs). Here we have investigated the advantages and capacities of using GPUs for cancer treatment planning based on a p...
With the arrival of “big data” (microarrays, screens, next-generation sequencing) in the life sci... more With the arrival of “big data” (microarrays, screens, next-generation sequencing) in the life sciences, standard analyses of these data for regular users of R now run into severe issues of computation time or computer memory. Many projects (including parallelisation efforts of the R core) offer R packages and functions that allow programming of solutions for large-scale analysis problems. However, these usually require familiarity with HPC programming as well as sufficient and funded time to employ, which is feasible for one-off analysis problems but impractical for common analysis methods.
One of the main hurdles of a broad distribution of PGAS approaches is the prevalence of MPI, whic... more One of the main hurdles of a broad distribution of PGAS approaches is the prevalence of MPI, which as a de-facto standard appears in the code basis of many applications. To take advantage of the PGAS APIs like GASPI without a major change in the code basis, interoperability between MPI and PGAS approaches needs to be ensured. In this article, we address this challenge by providing our study and preliminary performance results regarding interoperating GASPI and MPI on the performance crucial parts of the Ludwig and iPIC3D applications. In addition, we draw a strategy for better coupling of both APIs.
Accelerators and, in particular, Graphics Processing Units (GPUs) have emerged as promising compu... more Accelerators and, in particular, Graphics Processing Units (GPUs) have emerged as promising computing technologies which may be suitable for the future Exascale systems. Here, we present performance results of NekBone, a benchmark of the Nek5000 code, implemented with optimized OpenACC directives and GPUDirect communications. Nek5000 is a computational fluid dynamics code based on the spectral element method used for the simulation of incompressible flow. Results of an optimized NekBone version lead to 78 Gflops performance on a single node. In addition, a performance result of 609 Tflops has been reached on 16, 384 GPUs of the Titan supercomputer at Oak Ridge National Laboratory.
Coarrays are a native Fortran means for SPMD parallel programming, implementing a single-sided co... more Coarrays are a native Fortran means for SPMD parallel programming, implementing a single-sided communications model. Coarray Fortran belongs to the PGAS class of parallel languages. Relatively few tools support profiling of coarray programs. Profiling and tracing of pure coarray programs and mixed coarray+MPI miniapps was done in this work on 2 platforms. TAU (Tuning and Analysis Utilities) toolkit 2.25.2 was used with the Intel compiler 16.0.2 and the Intel MPI libraries 4.1.0 on Red Hat Enterprise Linux 6 cluster. CrayPAT (Performance Analysis Tool) was used on Cray XC30 system with CCE (Cray Compiler Environment) 8.4.1. Four coarray programs, of progressively increasing complexity, are studied. TAU profiling results on 32 cores and CrayPAT profiling data on 7k cores show that our hybrid MPI+coarray codes are relatively well balanced. We show that the Intel compiler maps coarray remote operations and synchronisation routines onto MPI-2 RMA calls. For the Intel implementation we sh...
We have developed miniapps from MPI finite element library ParaFEM and Fortran 2008 coarray cellu... more We have developed miniapps from MPI finite element library ParaFEM and Fortran 2008 coarray cellular automata library CGPACK. The miniapps represent multiscale fracture models of polycrystalline solids. The software from which these miniapps have been derived will improve predictive modelling in the automotive, aerospace, power generation, defense and manufacturing sectors. The libraries and miniapps are distributed under BSD license, so these can be used by computer scientists and hardware vendors to test various tools including compilers and performance monitoring applications. CrayPAT tools have been used for sampling and tracing analysis of the miniapps. Two routines with all-to-all communication structures have been identified a primary candidates for optimisation. New routines have been written implementing the nearest neighbour algorithm and using coarray collectives. Scaling limit for miniapps has been increased by a factor of 3, from about 2k to over 7k cores. The miniapps ...
Uploads
Papers by Luis Cebamanos