Search

In the world of high-performance computing, it is known that it can be beneficial to leave a small number of CPU cores unused, a practice termed undersubscription. However, undersubscription is rarely implemented in scientific applications of high-performance computing. We demonstrate the importance of calibrated undersubscription in Computational Fluid Dynamics simulations through the aggregated results of 1844 benchmarks. These benchmarks measured three hardware configurations and five different CFD models. On average, performance increased by 14% (weighted by node count). Performance improvements were most significant at large node counts, particularly when nearing a regime of negative scalability. We found that undersubscription could increase maximum performance by up to 50%; this advantage diminished as node count decreased but remained as much as 13% with a single node. In some cases, maximum performance was achieved with large numbers of free cores—nearly half of the cores in one case. Producing a regression from our dataset, we universally predict the optimal number of free cores as a function of cells per core. This regression achieves a 15% speed increase on average (again weighted by node count). [ABSTRACT FROM AUTHOR]

Cloud computing platforms have been continuously evolving. Features such as the Elastic Fabric Adapter (EFA) in the Amazon Web Services (AWS) platform have brought yet another revolution in the High Performance Computing (HPC) world, further accelerating the convergence of HPC and cloud computing. Other public clouds also support similar features further fueling this change. In this paper, we show how and why the performance of a large-scale computational fluid dynamics (CFD) HPC application on AWS competes very closely with the one on Beskow—a Cray XC40 supercomputer at the PDC Center for High-Performance Computing - in terms of cost-efficiency with strong scaling up to 2304 processes. We perform an extensive set of micro and macro benchmarks in both environments and conduct a comparative analysis. Until as recently as 2020 these benchmarks have notoriously yielded unsatisfactory results for the cloud platforms compared with on-premise infrastructures. Our aim is to access the HPC capabilities of the cloud, and in general to demonstrate how researchers can scale and evaluate the performance of their application in the cloud. [ABSTRACT FROM AUTHOR]

HPC or super-computing clusters are designed for executing computationally intensive operations that typically involve large scale I/O operations. This most commonly involves using a standard MPI library implemented in C/C++. The MPI-I/O performance in HPC clusters tends to vary significantly over a range of configuration parameters that are generally not taken into account by the algorithm. It is commonly left to individual practitioners to optimise I/O on a case by case basis at code level. This can often lead to a range of unforeseen outcomes. The ExSeisDat utility is built on top of the native MPI-I/O library comprising of Parallel I/O and Workflow Libraries to process seismic data encapsulated in SEG-Y file format. The SEG-Y File data structure is complex in nature, due to the alternative arrangement of trace header and trace data. Its size scales to petabytes and the chances of I/O performance degradation are further increased by ExSeisDat. This research paper presents a novel study of the changing I/O performance in terms of bandwidth, with the use of parallel plots against various MPI-I/O, Lustre (Parallel) File System and SEG-Y File parameters. Another novel aspect of this research is the predictive modelling of MPI-I/O behaviour over SEG-Y File benchmarks using Artificial Neural Networks (ANNs). The accuracy ranges from 62.5% to 96.5% over the set of trained ANN models. The computed Mean Square Error (MSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) values further support the generalisation of the prediction models. This paper demonstrates that by using our ANNs prediction technique, the configurations can be tuned beforehand to avoid poor I/O performance. [ABSTRACT FROM AUTHOR]

The paper presents a supercomputer parallel implementation of a brain inspired model combining a Python module simulating a layer of retina ganglion cells and NEST Simulator for a layer of spike timing neurons of Lateral geniculate nucleus (LGN) in the brain. Since the Python module appeared to be the bottleneck in the developed hierarchical model of human visual system, the proposed two parallel implementations of that module were combined with NEST Simulator in a single module. Simulations of developed module on different number of nodes and varying number of parallel processes were investigated and compared with respect to their time consumption. [ABSTRACT FROM AUTHOR]

Data centers are increasing in size and complexity, and we need scalable approaches to support their automated analysis and control. Performance counters and power consumption are their key "vital signs". State-of-the-Art (SoA) monitoring systems provide built-in tools to collect performance measurements, and custom solutions to get insight on their power consumption. However, with the increase in measurement resolution (in time and space) and the ensuing huge amount of measurement data to handle, new challenges arise, such as bottlenecks on the network bandwidth, storage and software overhead on the monitoring units. To face these challenges we propose a novel monitoring platform for data centers, which enables real-time high-resolution profiling (i.e., all available performance counters and the entire signal bandwidth of the power consumption at the plug—sampling up to 20 μ s —with an error below 1%) and analytics, both at the edge (node-level analysis) and on a centralized unit (cluster-level analysis). The monitoring infrastructure is completely out-of-band, scalable, technology agnostic and low cost, and it is already installed in a SoA high-performance compute cluster (i.e., D.A.V.I.D.E. —18th in Green500 November 2017). [ABSTRACT FROM AUTHOR]

Different domains of research are moving to cloud computing whether to carry out compute intensive experiments or to store large datasets. Large-scale computation in geophysical exploration is often inefficient, especially in the Just-in-time (JIT) environment. To alleviate this, we devised a new parallel magnetotelluric inversion method on high performance computing (HPC) multi-threads workloads cluster. This parallel algorithm adapted to single CPU or PC clusters with multi-threads workloads allocates different waves to each thread in a coarse-gained mode. In all multi-threads, the master thread deals with all parallel tasks, and other slave threads compute the electromagnetic field values of each wave in a parallel fork-join model. Experiments show that the proposed parallel algorithm not only achieves effective data accuracy, but is more efficient than the serial version. [ABSTRACT FROM AUTHOR]

Nowadays, many data centers use virtual machines (VMs) in order to achieve a more efficient use of hardware resources. The use of VMs provides a reduction in equipment and maintenance expenses as well as a lower electricity consumption. Nevertheless, current virtualization solutions, such as Xen, do not easily provide graphics processing units (GPUs) to applications running in the virtualized domain with the flexibility usually required in data centers (i.e., managing virtual GPU instances and concurrently sharing them among several VMs). Therefore, the execution of GPU-accelerated applications within VMs is hindered by this lack of flexibility. In this regard, remote GPU virtualization solutions may address this concern. In this paper we analyze the use of the remote GPU virtualization mechanism to accelerate scientific applications running inside Xen VMs. We conduct our study with six different applications, namely CUDA-MEME, CUDASW++, GPU-BLAST, LAMMPS, a triangle count application, referred to as TRICO, and a synthetic benchmark used to emulate different application behaviors. Our experiments show that the use of remote GPU virtualization is a feasible approach to address the current concerns of sharing GPUs among several VMs, featuring a very low overhead if an InfiniBand fabric is already present in the cluster. [ABSTRACT FROM AUTHOR]

In the past few years an online simulation service platform (named EDISON) has been applauded by several computational science and engineering communities in several countries. Though armed with multiple computing clusters and high-end storage resources, the platform has suffered from handling a huge amount of CPU-/IO-bound simulations that are most duplicated. Such intense simulations are normally admitted with no duplicate elimination and thus can adversely affect the performance of the platform. To address this performance concern, we propose a novel system, termed SUPERMAN, to seamlessly record and retrieve the provenances of previously executed simulations, and so prevent users from initiating duplicate and/or similar simulations using the limited computing resources. The system collects the simulation provenances based on a variant of a de-facto standard form, thereby offering interoperability. Based on the stored provenances, the system can provide useful simulation run statistics for users that need assistance. SUPERMAN also applies a hash-based duplicate elimination technique, resulting in making more efficient simulations on the platform. Finally, we show that the proposed proposed system could remove slightly over half of duplicate simulations on a variety of simulation software while obtaining about overall elapsed time savings of 30% and queuing time savings of 25%. [ABSTRACT FROM AUTHOR]

In high performance computing (HPC) resources’ extensive experiments are frequently executed. HPC resources (e.g. computing machines and switches) should be able to handle running several experiments in parallel. Typically HPC utilizes parallelization in programs, processing and data. The underlying network is seen as the only non-parallelized HPC component (i.e. no dynamic virtual slicing based on HPC jobs). In this scope we present an approach in this paper to utilize software defined networking (SDN) to parallelize HPC clusters among the different running experiments. We propose to accomplish this through two major components: A passive module (network mapper/remapper) to select for each experiment as soon as it starts the least busy resources in the network, and an SDN-HPC active load balancer to perform more complex and intelligent operations. Active load balancer can logically divide the network based on experiments’ host files. The goal is to reduce traffic to unnecessary hosts or ports. An HPC experiment should multicast, rather than broadcast to only cluster nodes that are used by the experiment. We use virtual tenant network modules in Opendaylight controller to create VLANs based on HPC experiments. In each HPC host, virtual interfaces are created to isolate traffic from the different experiments. The traffic between the different physical hosts that belong to the same experiment can be distinguished based on the VLAN ID assigned to each experiment. We evaluate the new approach using several HPC public benchmarks. Results show a significant enhancement in experiments’ performance especially when HPC cluster experiences running several heavy load experiments simultaneously. Results show also that this multi-casting approach can significantly reduce casting overhead that is caused by using a single cast for all resources in the HPC cluster. In comparison with InfiniBand networks that offer interconnect services with low latency and high bandwidth, HPC services based on SDN can provide two distinguished objectives that may not be possible with InfiniBand: The first objective is the integration of HPC with Ethernet enterprise networks and hence expanding HPC usage to much wider domains. The second objective is the ability to enable users and their applications to customize HPC services with different QoS requirements that fit the different needs of those applications and optimize the usage of HPC clusters. [ABSTRACT FROM AUTHOR]

A collection of virtual machines (VMs) interconnected with an overlay network with a layer 2 abstraction has proven to be a powerful, unifying abstraction for adaptive distributed and parallel computing on loosely-coupled environments. It is now feasible to allow VMs hosting high performance computing (HPC) applications to seamlessly bridge distributed cloud resources and tightly-coupled supercomputing and cluster resources. However, to achieve the application performance that the tightly-coupled resources are capable of, it is important that the overlay network not introduce significant overhead relative to the native hardware, which is not the case for current user-level tools, including our own existing VNET/U system. In response, we describe the design, implementation, and evaluation of a virtual networking system that has negligible latency and bandwidth overheads in 1-10 Gbps networks. Our system, VNET/P, is directly embedded into our publicly available Palacios virtual machine monitor (VMM). VNET/P achieves native performance on 1 Gbps Ethernet networks and very high performance on 10 Gbps Ethernet networks. The NAS benchmarks generally achieve over 95 % of their native performance on both 1 and 10 Gbps. We have further demonstrated that VNET/P can operate successfully over more specialized tightly-coupled networks, such as Infiniband and Cray Gemini. Our results suggest it is feasible to extend a software-based overlay network designed for computing at wide-area scales into tightly-coupled environments. [ABSTRACT FROM AUTHOR]

High performance cloud computing is behind the scene powering 'the next big thing' as the mainstream accelerator for innovation in many areas. We describe here how to accelerate inexpensive ARM-based computing nodes with high-end GPGPUs hosted on x86_64 machines using the GVirtuS general-purpose virtualization service. We draw the vision of a possible next generation computing clusters characterized by highly heterogeneous parallelism heading to a lower electric power demanding, less heat producing and more environmental friendliness. Preliminary but promising performance data suggest that this solution could be considered as part of the foundations of the next generation of high performance cloud computing components. [ABSTRACT FROM AUTHOR]

In December 2009 the 768-bit, 232-digit number RSA-768 was factored using the number field sieve. Overall, the computational challenge would take more than 1700 years on a single, standard core. In the article we present the heterogeneous computing approach, involving different compute clusters and Grid computing environments, used to solve this problem. [ABSTRACT FROM AUTHOR]