Sabela Ramos

Java in the High Performance Computing arena: Research, practice and experience

The rising interest in Java for High Performance Computing (HPC) is based on the appealing features of this language for programming multi-core cluster architectures, particularly the built-in networking and multithreading support, and the continuous increase in Java Virtual Machine (JVM) performance. However, its adoption in this area is being delayed by the lack of analysis of the existing programming options in Java for HPC and thorough and up-to-date evaluations of their performance, as well as the unawareness on current research projects in this field, whose solutions are needed in order to boost the embracement of Java in HPC. This paper analyzes the current state of Java for HPC, both for shared and distributed memory programming, presents related research projects, and finally, evaluates the performance of current Java HPC solutions and research developments on two shared memory environments and two InfiniBand multi-core clusters. The main conclusions are that: (1) the significant interest in Java for HPC has led to the development of numerous projects, although usually quite modest, which may have prevented a higher development of Java in this field; (2) Java can achieve almost similar performance to natively compiled languages, both for sequential and parallel applications, being an alternative for HPC programming; (3) the recent advances in the efficient support of Java communications on shared memory and low-latency networks are bridging the gap between Java and natively compiled applications in HPC. Thus, the good prospects of Java in this area are attracting the attention of both industry and academia, which can take significant advantage of Java adoption in HPC.

Performance analysis of HPC applications in the cloud

The scalability of High Performance Computing (HPC) applications depends heavily on the efficient support of network communications in virtualized environments. However, Infrastructure as a Service (IaaS) providers are more focused on deploying systems with higher computational power interconnected via high-speed networks rather than improving the scalability of the communication middleware. This paper analyzes the main performance bottlenecks in HPC application scalability on the Amazon EC2 Cluster Compute platform: (1) evaluating the communication performance on shared memory and a virtualized 10 Gigabit Ethernet network; (2) assessing the scalability of representative HPC codes, the NAS Parallel Benchmarks, using an important number of cores, up to 512; (3) analyzing the new cluster instances (CC2), both in terms of single instance performance, scalability and cost-efficiency of its use; (4) suggesting techniques for reducing the impact of the virtualization overhead in the scalability of communication-intensive HPC codes, such as the direct access of the Virtual Machine to the network and reducing the number of processes per instance; and (5) proposing the combination of message-passing with multithreading as the most scalable and cost-effective option for running HPC applications on the Amazon EC2 Cluster Compute platform. Highlights? Performance results of HPC applications in the cloud using up to 512 cores. ? Up-to-date performance evaluation of the Amazon EC2 Cluster Compute platform. ? High Performance Cloud Computing applications rely on scalable communication. ? Proposal of new techniques for increasing scalability of HPC codes in the cloud. ? Using several levels of parallelism is key for HPC scalability in the cloud.

Modeling communication in cache-coherent SMP systems: a case-study with Xeon Phi

Most multi-core and some many-core processors implement cache coherency protocols that heavily complicate the design of optimal parallel algorithms. Communication is performed implicitly by cache line transfers between cores, complicating the understanding of performance properties. We developed an intuitive performance model for cache-coherent architectures and demonstrate its use with the currently most scalable cache-coherent many-core architecture, Intel Xeon Phi. Using our model, we develop several optimal and optimized algorithms for complex parallel data exchanges. All algorithms that were developed with the model beat the performance of the highly-tuned vendor-specific Intel OpenMP and MPI libraries by up to a factor of 4.3. The model can be simplified to satisfy the tradeoff between complexity of algorithm design and accuracy. We expect that our model can serve as a vehicle for advanced algorithm design.

General‐purpose computation on GPUs for high performance cloud computing

Cloud computing is offering new approaches for High Performance Computing (HPC) as it provides dynamically scalable resources as a service over the Internet. In addition, General‐Purpose computation on Graphical Processing Units (GPGPU) has gained much attention from scientific computing in multiple domains, thus becoming an important programming model in HPC. Compute Unified Device Architecture (CUDA) has been established as a popular programming model for GPGPUs, removing the need for using the graphics APIs for computing applications. Open Computing Language (OpenCL) is an emerging alternative not only for GPGPU but also for any parallel architecture. GPU clusters, usually programmed with a hybrid parallel paradigm mixing Message Passing Interface (MPI) with CUDA/OpenCL, are currently gaining high popularity. Therefore, cloud providers are deploying clusters with multiple GPUs per node and high‐speed network interconnects in order to make them a feasible option for HPC as a Service (HPCaaS). This paper evaluates GPGPU for high performance cloud computing on a public cloud computing infrastructure, Amazon EC2 Cluster GPU Instances (CGI), equipped with NVIDIA Tesla GPUs and a 10 Gigabit Ethernet network. The analysis of the results, obtained using up to 64 GPUs and 256‐processor cores, has shown that GPGPU is a viable option for high performance cloud computing despite the significant impact that virtualized environments still have on network overhead, which still hampers the adoption of GPGPU communication‐intensive applications. Copyright

Analysis of I/O Performance on an Amazon EC2 Cluster Compute and High I/O Platform

Cloud computing is currently being explored by the scientific community to assess its suitability for High Performance Computing (HPC) environments. In this novel paradigm, compute and storage resources, as well as applications, can be dynamically provisioned on a pay-per-use basis. This paper presents a thorough evaluation of the I/O storage subsystem using the Amazon EC2 Cluster Compute platform and the recent High I/O instance type, to determine its suitability for I/O-intensive applications. The evaluation has been carried out at different layers using representative benchmarks in order to evaluate the low-level cloud storage devices available in Amazon EC2, ephemeral disks and Elastic Block Store (EBS) volumes, both on local and distributed file systems. In addition, several I/O interfaces (POSIX, MPI-IO and HDF5) commonly used by scientific workloads have also been assessed. Furthermore, the scalability of a representative parallel I/O code has also been analyzed at the application level, taking into account both performance and cost metrics. The analysis of the experimental results has shown that available cloud storage devices can have different performance characteristics and usage constraints. Our comprehensive evaluation can help scientists to increase significantly (up to several times) the performance of I/O-intensive applications in Amazon EC2 cloud. An example of optimal configuration that can maximize I/O performance in this cloud is the use of a RAID 0 of 2 ephemeral disks, TCP with 9,000 bytes MTU, NFS async and MPI-IO on the High I/O instance type, which provides ephemeral disks backed by Solid State Drive (SSD) technology.

Evaluation of Java for General Purpose GPU Computing

The presence of many-core units as accelerators has been increasing due to their ability to improve the performance of highly parallel workloads. General Purpose GPU(GPGPU) computing has allowed the graphical units to emerge as successful co-processors that can be employed to improve the performance of many different non-graphical applications with high parallel requirements, which make them suitable for many High Performance Computing workloads. While the main libraries developed to exploit the massive parallel capacity of GPUs are oriented to C/C++ programmers, there have been several efforts to extend this support to other languages. Among them, Java stands out for being one of the most extended languages and there are multiple projects that try to enable Java to take advantage of GPGPU computing. In this scenario, this paper presents an evaluation of the most relevant among the current solutions that exploit GPGPU computing in Java.

Capability Models for Manycore Memory Systems: A Case-Study with Xeon Phi KNL

Increasingly complex memory systems and onchip interconnects are developed to mitigate the data movement bottlenecks in manycore processors. One example of such a complex system is the Xeon Phi KNL CPU with three different types of memory, fifteen memory configuration options, and a complex on-chip mesh network connecting up to 72 cores. Users require a detailed understanding of the performance characteristics of the different options to utilize the system efficiently. Unfortunately, peak performance is rarely achievable and achievable performance is hardly documented. We address this with capability models of the memory subsystem, derived by systematic measurements, to guide users to navigate the complex optimization space. As a case study, we provide an extensive model of all memory configuration options for Xeon Phi KNL. We demonstrate how our capability model can be used to automatically derive new close-to-optimal algorithms for various communication functions yielding improvements 5x and 24x over Intel’s tuned OpenMP and MPI implementations, respectively. Furthermore, we demonstrate how to use the models to assess how efficiently a bitonic sort application utilizes the memory resources. Interestingly, our capability models predict and explain that the high bandwidthMCDRAM does not improve the bitonic sort performance over DRAM.

Evaluation of messaging middleware for high-performance cloud computing

Cloud computing is posing several challenges, such as security, fault tolerance, access interface singularity, and network constraints, both in terms of latency and bandwidth. In this scenario, the performance of communications depends both on the network fabric and its efficient support in virtualized environments, which ultimately determines the overall system performance. To solve the current network constraints in cloud services, their providers are deploying high-speed networks, such as 10 Gigabit Ethernet. This paper presents an evaluation of high-performance computing message-passing middleware on a cloud computing infrastructure, Amazon EC2 cluster compute instances, equipped with 10 Gigabit Ethernet. The analysis of the experimental results, confronted with a similar testbed, has shown the significant impact that virtualized environments still have on communication performance, which demands more efficient communication middleware support to get over the current cloud network limitations.

Design of efficient Java message-passing collectives on multi-core clusters

This paper presents a scalable and efficient Message-Passing in Java (MPJ) collective communication library for parallel computing on multi-core architectures. The continuous increase in the number of cores per processor underscores the need for scalable parallel solutions. Moreover, current system deployments are usually multi-core clusters, a hybrid shared/distributed memory architecture which increases the complexity of communication protocols. Here, Java represents an attractive choice for the development of communication middleware for these systems, as it provides built-in networking and multithreading support. As the gap between Java and compiled languages performance has been narrowing for the last years, Java is an emerging option for High Performance Computing (HPC).Our MPJ collective communication library increases Java HPC applications performance on multi-core clusters: (1) providing multi-core aware collective primitives; (2) implementing several algorithms (up to six) per collective operation, whereas publicly available MPJ libraries are usually restricted to one algorithm; (3) analyzing the efficiency of thread-based collective operations; (4) selecting at runtime the most efficient algorithm depending on the specific multi-core system architecture, and the number of cores and message length involved in the collective operation; (5) supporting the automatic performance tuning of the collectives depending on the system and communication parameters; and (6) allowing its integration in any MPJ implementation as it is based on MPJ point-to-point primitives. A performance evaluation on an InfiniBand and Gigabit Ethernet multi-core cluster has shown that the implemented collectives significantly outperform the original ones, as well as higher speedups when analyzing the impact of their use on collective communications intensive Java HPC applications. Finally, the presented library has been successfully integrated in MPJ Express (http://mpj-express.org), and will be distributed with the next release.

Multithreaded and Spark parallelization of feature selection filters

Abstract Vast amounts of data are generated every day, constituting a volume that is challenging to analyze. Techniques such as feature selection are advisable when tackling large datasets. Among the tools that provide this functionality, Weka is one of the most popular ones, although the implementations it provides struggle when processing large datasets, requiring excessive times to be practical. Parallel processing can help alleviate this problem, effectively allowing users to work with Big Data. The computational power of multicore machines can be harnessed by using multithreading and distributed programming, effectively helping to tackle larger problems. Both these techniques can dramatically speed up the feature selection process allowing users to work with larger datasets. The reimplementation of four popular feature selection algorithms included in Weka is the focus of this work. Multithreaded implementations previously not included in Weka as well as parallel Spark implementations were developed for each algorithm. Experimental results obtained from tests on real-world datasets show that the new versions offer significant reductions in processing times.