An Approach to Formalization and Analysis of Message Passing Libraries

High performance computing (HPC) seems to be one of the last monopolies of low-level languages like C and FORTRAN. The defacto standard for HPC, the Message Passing Interface (MPI), defines APIs for C, FORTRAN and C++ only. This paper evaluates current alternatives among interpreted languages, specifically Python and C#. MPI library wrappers for both languages are examined and their performance is compared to native (C) OpenMPI using two benchmarks. Both languages compare favorably in code and performance effectiveness.

Understanding the behavior of parallel applications that use the Message Passing Interface (MPI) is critical for optimizing communication performance. Performance tools for MPI currently rely on the PMPI Profiling Interface or the MPI Tools Information Interface, MPI_T, for portably collecting information for performance measurement and analysis. While tools using these interfaces have proven to be extremely valuable for performance tuning, these interfaces only provide synchronous information, i.e., when an MPI or an MPI_T function is called. There is currently no option for collecting information about asynchronous events from within the MPI library. In this work we propose a callback-driven interface for event notification from MPI implementations. Our approach is integrated in the existing MPI_T interface and provides a portable API for tools to discover and register for events of interest. We demonstrate the functionality and usability of the interface with a prototype implementation in Open MPI, a small logging tool (MEL) and the measurement infrastructure Score-P.

This paper describes an implementation of the Message Passing Interface (MPI) on the Portals 3.0 data movement layer. Portals 3.0 provides low-level building blocks that are flexible enough to support higher-level message passing layers, such as MPI, very efficiently. Portals 3.0 is also designed to allow for programmable network interface cards to offload message processing from the host processor, allowing for the ability to overlap computation and MPI communication. We describe the basic building blocks in Portals 3.0, show how they can be put together to implement MPI, and describe the protocols of our MPI implementation. We look at several key operations within the implementation and describe the effects that a Portals 3.0 implementation has on scalability and performance. We also present preliminary performance results from our implementation for Myrinet.

High Performance Computing (HPC) applications commonly use Message Passing Interface (MPI) to scale complex workloads on distributed memory architectures, while Graphical User Interface (GUI) is the primary basis for pre-processing and post-processing. Containerization enables portability and reproducibility for applications. Docker is a widely used containerization platform but cannot be easily adopted in secure HPC environments due to the need for root privileges to run Docker containers and the ensuing security issues, as well as the lack of native support for Message Passing Interface (MPI) and HPC workload managers. In addition to offering portability and reproducibility to scientific computing, Singularity and Sarus are container runtimes that enable containers to run on HPC with user credentials by removing security concerns and allowing reuse of Docker containers without any changes. This work presents approaches to creating and running MPI and GUI applications in an isolated container environment on HPC in the cloud. We provide performance evaluation of application containers transformed from Docker to Singularity and Sarus in the context of native on HPC in the cloud.

In prevailing technology trends it is apparent that the network requirements and technologies will advance in future. Therefore the need of High Performance Computing (HPC) based implementation for interconnecting clusters is comprehensible for scalability of clusters. Grid computing provides global infrastructure of interconnecting clusters consisting of dispersed computing resources over Internet. On the other hand the leading model for HPC programming is Message Passing Interface (MPI). As compared to Grid computing, MPI is better suited for solving most of the complex computational problems. MPI itself is restricted to a single cluster. It does not support message passing over the internet to use the computing resources of different clusters in an optimal way. We propose a model that provides message passing capabilities between parallel applications over the internet. The proposed model is based on Architecture for Java Universal Message Passing (A-JUMP) framework and Enterprise Service Bus (ESB) named as High Performance Computing Bus. The HPC Bus is built using ActiveMQ. HPC Bus is responsible for communication and message passing in an asynchronous manner. Asynchronous mode of communication offers an assurance for message delivery as well as a fault tolerance mechanism for message passing. The idea presented in this paper effectively utilizes wide-area intercluster networks. It also provides scheduling, dynamic resource discovery and allocation, and sub-clustering of resources for different jobs. Performance analysis and comparison study of the proposed framework with P2P-MPI are also presented in this paper.

Process failure is projected to become a normal event for many long running and scalable High Performance Computing (HPC) applications. As such many application developers are investigating Algorithm Based Fault Tolerance (ABFT) techniques to improve the efficiency of application recovery beyond what existing checkpoint/restart techniques alone can provide. Unfortunately for these application developers the libraries that their applications depend upon, like Message Passing Interface (MPI), do not have standardized fault tolerance semantics. This paper introduces the reader to a set of run-through stabilization semantics being developed by the MPI Forum's Fault Tolerance Working Group to support ABFT. Using a well-known ring communication program as the running example, this paper illustrates to application developers new to ABFT some of the issues that arise when designing a fault tolerant application. The ring program allows the paper to focus on the communication-level issues rather than the data preservation mechanisms covered by existing literature. This paper highlights a common set of issues that application developers must address in their design including program control management, duplicate message detection, termination detection, and testing. The discussion provides application developers new to ABFT with an introduction to both new interfaces becoming available, and a range of design issues that they will likely need to address regardless of their research domain.

According to Moore's Law, computer speeds are expected to double approximately every 2 years. But with the current challenges that computer manufacturers are facing to double speeds of individual processors, due to various reasons, i.e., processor temperatures; multiprocessor architectures have become more popular nowadays. Eventually, this has led to an increased interest in standards for writing parallel applications. The Message Passing Interface (MPI) has become the de facto standard for writing applications to run on such systems. Our growing needs for the latest high performance computing solutions in Saudi Aramco, the world's largest oil producing company, has given us the opportunity to evaluate three of the most commonly used MPI implementations, MVAPICH2, Open MPI, and Intel MPI on Intel's latest Nehalem processor. In this paper, we describe our test bed environment along with the evaluations that we did using some of the Intel MPI Benchmarks (IMB). We discuss the results in terms of bandwidth speed, execution time, and scalability when running on up to 512 processors on an Infiniband Nehalem Linux cluster with 64 nodes using the three implementations of MPI. We finally state our conclusions, recommendations, and future work.

Chapter 8 - The Essential MPI

Various benchmarks exist for assessing performance characteristics of commodity hardware utilized in high performance computing (HPC) cluster environments. An additional assessment of bandwidth for Intel's 44-core processors considers network saturation via message passing focused on indirectly connected cores. The benchmark developed in this work provides a means of measuring bandwidth of two Broadwell nodes when communicating inter-chassis with all cores of each node utilized. This benchmark was developed in three phases using Message Passing Interface (MPI). First, the bandwidth measure was tested for MPI_Send operations in point-to-point communication. Second, the merge sort algorithm was implemented as a means of assessing bandwidth and a tree-structured communication algorithm was developed and implemented to maximize internode communication and minimize intra-node communication. Third, the benchmark was tuned to the configuration of the Distributed Environment for Academic Computing (DEAC) Cluster. A second benchmark removing local computation from the merge sort algorithm effectively performs a gather operation in a tree-structure. These benchmarks were tested and compared with relevant Intel MPI Benchmarks (IMB) tests. In the developed benchmarks, the nodes requested significantly more bandwidth than in the existing MPI_Gather operation and IMB benchmarks due to the large size of the messages simultaneously placed on the network.

MCM: A new MPI Communication Management for Cloud Environments

The Message Passing Interface (MPI) remains the dominant programming model for scientific applications running on today’s high-performance computing (HPC) systems. This dominance stems from MPI’s powerful semantics for inter-process communication that has enabled scientists to write applications for simulating important physical phenomena. MPI does not, however, specify how messages and synchronization should be carried out. Those details are typically dependent on low-level architecture details and the message characteristics of the application. Therefore, analyzing an applications MPI usage is critical to tuning MPI’s performance on a particular platform. The results of this analysis is typically a discussion of average message sizes for a workload or set of workloads. While a discussion of the message average might be the most intuitive summary statistic, it might not be the most useful in terms of representing the entire message size dataset for an application. Using a previously developed MPI trace collector, we analyze the MPI message traces for a number of key MPI workloads. Through this analysis, we demonstrate that the average, while easy and efficient to calculate, may not be a good representation of all subsets of application messages sizes, with median and mode of message sizes being a superior choice in most cases. We show that the problem with using the average relate to the multi-modal nature of the distribution of point-to-point messages. Finally, we show that while scaling a workload has little discernible impact on which measures of central tendency are representative of the underlying data, different input descriptions can significantly impact which metric is most effective. The results and analysis in this paper have the potential for providing valuable guidance on how we as a community should discuss and analyze MPI message data for scientific applications.

Parallel and distributed computing have enabled development of much more scalable software. However, developing concurrent software requires the programmer to be aware of nondeterminism, data races, and deadlocks. MPI (message passing interface) is a popular standard for writing message-oriented distributed applications. Some messages in MPI systems can be processed by one of the many machines and in many possible orders. This non-determinism can affect the result of an MPI application. The alternate results may or may not be correct. To verify MPI applications, we need to check all these possible orderings and use an application specific oracle to decide if these orderings give correct output.MPJ Express is an open source Java implementation of the MPI standard. Model checking of MPI Java programs is a challenging task due to their parallel nature. We developed a Java based model of MPJ Express, where processes are modeled as threads, and which can run unmodified MPI Java programs on a single system. This model enabled us to adapt the Java PathFinder explicit state software model checker (JPF) using a custom listener to verify our model running real MPI Java programs. The evaluation of our approach shows that model checking reveals incorrect system behavior that results in very intricate message orderings.

Computational modelling in material science involves mathematical abstractions of force fields between particles with the aim to postulate, develop and understand materials by simulation. The aggregated pairwise interactions of the material’s particles lead to a deduction of its macroscopic behaviours. For practically meaningful macroscopic scales, a large amount of data are generated, leading to vast execution times. Simulation times of hours, days or weeks for moderately sized problems are not uncommon. The reduction of simulation times, improved result accuracy and the associated software and hardware engineering challenges are the main motivations for many of the ongoing researches in the computational sciences. This contribution is concerned mainly with simulations that can be done on a “standalone” computer based on Message Passing Interfaces (MPI), parallel code running on hardware platforms with wide specifications, such as single/multi- processor, multi-core machines with minimal reconfiguration for upward scaling of computational power. The widely available, documented and standardized MPI library provides this functionality through the MPI_Comm_size (), MPI_Comm_rank () and MPI_Reduce () functions. A survey of the literature shows that relatively little is written with respect to the efficient extraction of the inherent computational power in a cluster. In this work, we discuss the main avenues available to tap into this extra power without compromising computational accuracy. We also present methods to overcome the high inertia encountered in single-node-based computational molecular dynamics. We begin by surveying the current state of the art and discuss what it takes to achieve parallelism, efficiency and enhanced computational accuracy through program threads and message passing interfaces. Several code illustrations are given. The pros and cons of writing raw code as opposed to using heuristic, third-party code are also discussed. The growing trend towards graphical processor units and virtual computing clouds for high-performance computing is also discussed. Finally, we present the comparative results of vacancy formation energy calculations using our own parallelized standalone code called Verlet–Stormer velocity (VSV) operating on 30,000 copper atoms. The code is based on the Sutton–Chen implementation of the Finnis–Sinclair pairwise embedded atom potential. A link to the code is also given.

There are several popular Big Data processing frameworks including Apache Spark, Dask, and Ray. The Apache Spark software provides an easy-to-use high-level API in different languages including Scala, Java, and Python. Spark supports parallel and distributed execution of user workloads by supporting communication using an event-driven framework called Netty. Some efforts - including RDMA-Spark and SparkUCX - were made in the past to optimize Apache Spark on High-Performance Computing (HPC) systems equipped with high-performance interconnects like InfiniBand. In the HPC community, Message Passing Interface (MPI) libraries are widely adopted for parallelizing science and engineering applications. This paper presents MPI4Spark which uses MPI for communication in a parallel and distributed setting on HPC systems. MPI4Spark can launch the Spark ecosystem using MPI launchers to utilize MPI communication inside the Big Data framework. It also maintains isolation for application execution on worker nodes by forking new processes using Dynamic Process Management (DPM). It bridges semantic differences between the event-driven communication in Spark compared to the application-driven communication engine in MPI. MPI4Spark also provides portability and performance benefits as it is capable of utilizing popular HPC interconnects including InfiniBand, Omni-Path, Slingshot, and others. The performance of MPI4Spark is evaluated against RDMA-Spark and Vanilla Spark using OSU HiBD Benchmarks (OHB) and Intel HiBench that contain a variety of Resilient Distributed Dataset (RDD), Graph Processing, and Machine Learning workloads. This evaluation is done on three HPC systems including TACC Frontera, TACC Stampede2, and an internal cluster. MPI4Spark outperforms Vanilla Spark and RDMA-Spark by 4.23x and 2.04x, respectively, on the TACC Frontera system using 448 processing cores (8 Spark workers) for the GroupByTest benchmark in OHB. The communication performance of MPI4Spark is 13.08x and 5.56x better than Vanilla Spark and RDMA-Spark, respectively.

Parallel and distributed deep learning (PDNN) has become an effective strategy to reduce the long training times of large-scale deep neural networks. Mainstream PDNN software packages based on the message-passing interface (MPI) and employing synchronous stochastic gradient descent rely crucially on the performance of MPI allreduce collective communication routine.In this work, we propose a novel scalable universal allreduce meta-algorithm called SUARA. In general, SUARA consists of L serial steps, where L≥2, executed by all MPI processes involved in the allreduce operation. At each step, SUARA partitions this set of processes into subsets, which execute optimally selected library allreduce algorithms to solve sub-allreduce problems on these subsets in parallel, to accomplish the whole allreduce operation after completing all the L steps. We then design, theoretically study and implement a two-step SUARA (L=2) called SUARA2 on top of the Open MPI library. We prove that the theoretical asymptotic speedup of SUARA2 executed by P processes over the base Open MPI routine is O(P). Our experiments on Shaheen-II supercomputer employing 1024 nodes demonstrate over 2x speedup of SUARA2 over native Open MPI allreduce routine, which translates into the performance improvement of training of ResNet-50 DNN on ImageNet by 9%.