Analytic, first-principles performance modeling of distributed-memory parallel codes is notoriously imprecise. Even for applications with extremely regular and homogeneous compute-communicate phases, simply adding communication time to computation time does often not yield a satisfactory prediction of parallel runtime due to deviations from the expected simple lockstep pattern caused by system noise, variations in communication time, and inherent load imbalance. In this paper, we highlight the specific cases of provoked and spontaneous desynchronization of memory-bound, bulk-synchronous pure MPI and hybrid MPI+OpenMP programs. Using simple microbenchmarks we observe that although desynchronization can introduce increased waiting time per process, it does not necessarily cause lower resource utilization but can lead to an increase in available bandwidth per core. In case of significant communication overhead, even natural noise can shove the system into a state of automatic overlap of communication and computation, improving the overall time to solution. The saturation point, i.e., the number of processes per memory domain required to achieve full memory bandwidth, is pivotal in the dynamics of this process and the emerging stable wave pattern. We also demonstrate how hybrid MPI-OpenMP programming can prevent desirable desynchronization by eliminating the bandwidth bottleneck among processes. A Chebyshev filter diagonalization application is used to demonstrate some of the observed effects in a realistic setting.
The performance of highly parallel applications on distributed-memory systems is influenced by many factors. Analytic performance modeling techniques aim to provide insight into performance limitations and are often the starting point of optimization efforts. However, coupling analytic models across the system hierarchy (socket, node, network) fails to encompass the intricate interplay between the program code and the hardware, especially when execution and communication bottlenecks are involved. In this paper we investigate the effect of "bottleneck evasion" and how it can lead to automatic o…
The hybrid MPI+X programming paradigm, where X refers to threads or GPUs, has gained prominence in the high-performance computing arena. This corresponds to a trend of system architectures growing more heterogeneous. The current MPI standard only specifies the compatibility levels between MPI and threading runtimes. No MPI concept or interface exists for applications to pass thread context or GPU stream context to MPI implementations explicitly. This lack has made performance optimization complicated in some cases and impossible in other cases. We propose a new concept in MPI, called MPIX stre…
MPI accomplishes portable, standardized message-passing between processes by exposing a standard API that hides the implementation of the underlying mechanism for message passing. Until now, checkpointing an MPI program required knowledge of these underlying mechanisms. Through the addition of a proxy, we demonstrate that MPI programs can be checkpointed and restarted regardless of the MPI implementation utilized. Further, proxies may enable MPI programs to be checkpointed on one MPI implementation, and restarted on another.
Modern interconnects offer remote direct memory access (RDMA) features. Yet, most applications rely on explicit message passing for communications albeit their unwanted overheads. The MPI-3.0 standard defines a programming interface for exploiting RDMA networks directly, however, it's scalability and practicability has to be demonstrated in practice. In this work, we develop scalable bufferless protocols that implement the MPI-3.0 specification. Our protocols support scaling to millions of cores with negligible memory consumption while providing highest performance and minimal overheads. To ar…
ClassdescMP is a distributed memory parallel programming system for use with C++ and MPI. It uses the Classdesc reflection system to ease the task of building complicated messages to be sent between processes. It doesn't hide the underlying MPI API, so it is an augmentation of MPI capabilities. Users can still call standard MPI function calls if needed for performance reasons.
The use of hybrid scheme combining the message passing programming models for inter-node parallelism and the shared memory programming models for node-level parallelism is widely spread. Existing extensive practices on hybrid Message Passing Interface (MPI) plus Open Multi-Processing (OpenMP) programming account for its popularity. Nevertheless, strong programming efforts are required to gain performance benefits from the MPI+OpenMP code. An emerging hybrid method that combines MPI and the MPI shared memory model (MPI+MPI) is promising. However, writing an efficient hybrid MPI+MPI program -- e…
A model of MPI synchronization communication programs is presented and its three basic simplified models are also defined. A series of theorems and methods for deciding whether deadlocks will occur among the three models are given and proved strictly. These theories and methods for simple models' deadlock detection are the necessary base for real MPI program deadlock detection. The methods are based on a static analysis through programs and with runtime detection in necessary cases and they are able to determine before compiling whether it will be deadlocked for two of the three basic models. …
Message Passing Interface (MPI) is the most commonly used paradigm in writing parallel programs since it can be employed not only within a single processing node but also across several connected ones. Data flow analysis concepts, techniques and tools are needed to understand and analyze MPI-based programs to detect bugs arise in these programs. In this paper we propose two automated techniques to analyze and debug MPI-based programs source codes.
In this paper, we present OMP2MPI a tool that generates automatically MPI source code from OpenMP. With this transformation the original program can be adapted to be able to exploit a larger number of processors by surpassing the limits of the node level on large HPC clusters. The transformation can also be useful to adapt the source code to execute in distributed memory many-cores with message passing support. In addition, the resulting MPI code can be used as an starting point that still can be further optimized by software engineers. The transformation process is focused on detecting OpenMP…
The current trend of multicore architectures on shared memory systems underscores the need of parallelism. While there are some programming model to express parallelism, thread programming model has become a standard to support these system such as OpenMP, and POSIX threads. MPI (Message Passing Interface) which remains the dominant model used in high-performance computing today faces this challenge. Previous version of MPI which is MPI-1 has no shared memory concept, and Current MPI version 2 which is MPI-2 has a limited support for shared memory systems. In this research, MPI-2 version of MP…
Message Passing Interfaces (MPI) plays an important role in parallel computing. Many parallel applications are implemented as MPI programs. The existing methods of bug detection for MPI programs have the shortage of providing both input and non-determinism coverage, leading to missed bugs. In this paper, we employ symbolic execution to ensure the input coverage, and propose an on-the-fly schedule algorithm to reduce the interleaving explorations for non-determinism coverage, while ensuring the soundness and completeness. We have implemented our approach as a tool, called MPISE, which can autom…
Parallel programs written using the standard Message Passing Interface (MPI) frequently depend upon the ability to efficiently execute collective operations. MPI_Scan is a collective operation defined in MPI that implements parallel prefix scan which is very useful primitive operation in several parallel applications. This operation can be very time consuming. In this paper, we explore the use of hardware programmable network interface cards utilizing standard media access protocols for offloading the MPI_Scan operation to the underlying network. Our work is based upon the NetFPGA - a programm…
MPI applications begin with a fixed number of rank and, by default, the rank remains constant throughout the application's lifetime. The developer can choose to increase the rank by dynamically spawning MPI processes. However doing this manually adds complexity to the MPI application. Making the MPI applications malleable \cite{b20} would allow HPC applications to have the same elasticity as that of cloud applications. We propose multiple approaches to change the rank of an MPI program agnostic to the modification of the user code. We use checkpointing as a tool to achieve mutability of rank b…
Clusters of SMP nodes provide support for a wide diversity of parallel programming paradigms. Combining both shared memory and message passing parallelizations within the same application, the hybrid MPI-OpenMP paradigm is an emerging trend for parallel programming to fully exploit distributed shared-memory architecture. In this paper, we improve the performance of MPEG-2 encoder and n-body simulation by employing the hybrid MPI-OpenMP programming paradigm on SMP clusters. The hierarchical image data structure of the MPEG bit-stream is eminently suitable for the hybrid model to achieve multipl…
Hybrid parallel programming models combining distributed and shared memory paradigms are well established in high-performance computing. The classical prototype of hybrid programming in HPC is MPI/OpenMP, but many other combinations are being investigated. Recently, the data-dependency driven, task parallel model for shared memory parallelisation named StarSs has been suggested for usage in combination with MPI. In this paper we apply hybrid MPI/StarSs to a Lattice-Boltzmann code. In particular, we present the hybrid programming model, the benefits we expect, the challenges in porting, and fin…
The Adapteva Epiphany many-core architecture comprises a 2D tiled mesh Network-on-Chip (NoC) of low-power RISC cores with minimal uncore functionality. It offers high computational energy efficiency for both integer and floating point calculations as well as parallel scalability. Yet despite the interesting architectural features, a compelling programming model has not been presented to date. This paper demonstrates an efficient parallel programming model for the Epiphany architecture based on the Message Passing Interface (MPI) standard. Using MPI exploits the similarities between the Epiphan…
Program correctness is one of the most difficult challenges in parallel programming. Message Passing Interface MPI is widely used in writing parallel applications. Since MPI is not a compiled language, the programmer will be enfaced with several programming bugs.This paper presents the most common programming bugs arise in MPI programs to help the programmer to compromise between the advantage of parallelism and the extra effort needed to detect and fix such bugs. An algebraic specification of an MPI-like programming language, called Simple MPI (SMPI), to be used in writing MPI programs specif…
Asynchronous programming models (APM) are gaining more and more traction, allowing applications to expose the available concurrency to a runtime system tasked with coordinating the execution. While MPI has long provided support for multi-threaded communication and non-blocking operations, it falls short of adequately supporting APMs as correctly and efficiently handling MPI communication in different models is still a challenge. Meanwhile, new low-level implementations of light-weight, cooperatively scheduled execution contexts (fibers, aka user-level threads (ULT)) are meant to serve as a bas…
We describe a program for the parallel implementation of multiple runs of XSTAR, a photoionization code that is used to predict the physical properties of an ionized gas from its emission and/or absorption lines. The parallelization program, called MPI_XSTAR, has been developed and implemented in the C++ language by using the Message Passing Interface (MPI) protocol, a conventional standard of parallel computing. We have benchmarked parallel multiprocessing executions of XSTAR, using MPI_XSTAR, against a serial execution of XSTAR, in terms of the parallelization speedup and the computing resou…
The Message Passing Interface (MPI) has been extremely successful as a portable way to program high-performance parallel computers. This success has occurred in spite of the view of many that message passing is difficult and that other approaches, including automatic parallelization and directive-based parallelism, are easier to use. This paper argues that MPI has succeeded because it addresses all of the important issues in providing a parallel programming model.
It is very common to use dynamic methods to detect deadlocks in MPI programs for the reason that static methods have some restrictions. To guarantee high reliability of some important MPI-based application software, a model of MPI synchronization communication is abstracted and a type of static method is devised to examine deadlocks in such modes. The model has three forms with different complexity: sequential model, single-loop model and nested-loop model. Sequential model is a base for all models. Single-loop model must be treated with a special type of equation group and nested-loop model e…
Detecting deadlocks in MPI synchronization communication programs is very difficult and need building program models. All complex models are based on sequential models. The sequential model is mapped into a set of character strings and its deadlock detection problem is translated into an equivalent multi-queue string matching problem. An algorithm is devised and implemented to statically detect deadlocks in sequential models of MPI synchronization communication programs. The time and space complexity of the algorithm is O(n) where n is the amount of message in model. The algorithm is better th…
Dynamic voltage and frequency scaling proves to be an efficient way of reducing energy consumption of servers. Energy savings are typically achieved by setting a well-chosen frequency during some program phases. However, determining suitable program phases and their associated optimal frequencies is a complex problem. Moreover, hardware is constrained by non negligible frequency transition latencies. Thus, various heuristics were proposed to determine and apply frequencies, but evaluating their efficiency remains an issue. In this paper, we translate the energy minimization problem into a mixe…
MPI-rical: Data-Driven MPI Distributed Parallelism Assistance with Transformers
Ben-Gurion University（本-古里安大学）；Israel Atomic Energy Commission（以色列原子能委员会）；Intel Labs（英特尔实验室）；Nuclear Research Center – Negev（内盖夫核研究中心）；Technion – Israel Institute of Technology（以色列理工学院）
MPI applications matter. However, with the advent of many-core processors, traditional MPI applications are challenged to achieve satisfactory performance. This is due to the inability of these applications to respond to load imbalances, to reduce serialization imposed by synchronous communication patterns, to overlap communication with computation and finally to deal with increasing memory overheads. The MPI specification provides asynchronous calls to mitigate some of these factors. However, application developers rarely make the effort to apply them efficiently. In this work, we present a m…
The Message Passing Interface (MPI) is the de facto standard message-passing infrastructure for developing parallel applications. Two decades after the first version of the library specification, MPI-based applications are nowadays routinely deployed on super and cluster computers. These applications, written in C or Fortran, exhibit intricate message passing behaviours, making it hard to statically verify important properties such as the absence of deadlocks. Our work builds on session types, a theory for describing protocols that provides for correct-by-construction guarantees in this regard…
A Partitioned Global Address Space (PGAS) approach treats a distributed system as if the memory were shared on a global level. Given such a global view on memory, the user may program applications very much like shared memory systems. This greatly simplifies the tasks of developing parallel applications, because no explicit communication has to be specified in the program for data exchange between different computing nodes. In this paper we present DART, a runtime environment, which implements the PGAS paradigm on large-scale high-performance computing clusters. A specific feature of our imple…
Composability is one of seven reasons for the long-standing and continuing success of MPI. Extending MPI by composing its operations with user-level operations provides useful integration with the progress engine and completion notification methods of MPI. However, the existing extensibility mechanism in MPI (generalized requests) is not widely utilized and has significant drawbacks. MPI can be generalized via scheduled communication primitives, for example, by utilizing implementation techniques from existing MPI-3 nonblocking collectives and from forthcoming MPI-4 persistent and partitioned …
Message Passing Interface (MPI) is widely used to implement parallel programs. Although Windowsbased architectures provide the facilities of parallel execution and multi-threading, little attention has been focused on using MPI on these platforms. In this paper we use the dual core Window-based platform to study the effect of parallel processes number and also the number of cores on the performance of three MPI parallel implementations for some sorting algorithms.
Lomonosov Moscow State University（莫斯科国立罗蒙诺索夫大学）；Applied Research Center for Computer Networks (ARCCN)（计算机网络应用研究中心）
The problem of MPI programs execution time prediction on a certain set of computer installations is considered. This problem emerges with orchestration and provisioning a virtual infrastructure in a cloud computing environment over a heterogeneous network of computer installations: supercomputers or clusters of servers (e.g. mini data centers). One of the key criteria for the effectiveness of the cloud computing environment is the time staying by the program inside the environment. This time consists of the waiting time in the queue and the execution time on the selected physical computer inst…
The Message Passing Interface (MPI) is the prevalent programming model used on today's supercomputers. Therefore, MPI library developers are looking for the best possible performance (shortest run-time) of individual MPI functions across many different supercomputer architectures. Several MPI benchmark suites have been developed to assess the performance of MPI implementations. Unfortunately, the outcome of these benchmarks is often neither reproducible nor statistically sound. To overcome these issues, we show which experimental factors have an impact on the run-time of blocking collective MP…
The Message Passing Interface (MPI) is the most commonly used application programming interface for process communication on current large-scale parallel systems. Due to the scale and complexity of modern parallel architectures, it is becoming increasingly difficult to optimize MPI libraries, as many factors can influence the communication performance. To assist MPI developers and users, we propose an automatic way to check whether MPI libraries respect self-consistent performance guidelines for collective communication operations. We introduce the PGMPI framework to detect violations of perfo…
The Message Passing Interface specification (MPI) defines a portable message-passing API used to program parallel computers. MPI programs manifest a number of challenges on what concerns correctness: sent and expected values in communications may not match, resulting in incorrect computations possibly leading to crashes; and programs may deadlock resulting in wasted resources. Existing tools are not completely satisfactory: model-checking does not scale with the number of processes; testing techniques wastes resources and are highly dependent on the quality of the test set. As an alternative, …
We discuss the use of both MPI and OpenMP in the teaching of senior undergraduate and junior graduate classes in parallel programming. We briefly introduce the OpenMP standard and discuss why we have chosen to use it in parallel programming classes. Advantages of using OpenMP over message passing methods are discussed. We also include a brief enumeration of some of the drawbacks of using OpenMP and how these drawbacks are being addressed by supplementing OpenMP with additional MPI codes and projects. Several projects given in my class are also described in this paper.
Hybrid MPI+threads programming is gaining prominence, but, in practice, applications perform slower with it compared to the MPI everywhere model. The most critical challenge to the parallel efficiency of MPI+threads applications is slow MPI_THREAD_MULTIPLE performance. MPI libraries have recently made significant strides on this front, but to exploit their capabilities, users must expose the communication parallelism in their MPI+threads applications. Recent studies show that MPI 4.0 provides users with new performance-oriented options to do so, but our evaluation of these new mechanisms shows…
The advent of multi-/many-core processors in clusters advocates hybrid parallel programming, which combines Message Passing Interface (MPI) for inter-node parallelism with a shared memory model for on-node parallelism. Compared to the traditional hybrid approach of MPI plus OpenMP, a new, but promising hybrid approach of MPI plus MPI-3 shared-memory extensions (MPI+MPI) is gaining attraction. We describe an algorithmic approach for collective operations (with allgather and broadcast as concrete examples) in the context of hybrid MPI+MPI, so as to minimize memory consumption and memory copies. …
Computationally-intensive loops are the primary source of parallelism in scientific applications. Such loops are often irregular and a balanced execution of their loop iterations is critical for achieving high performance. However, several factors may lead to an imbalanced load execution, such as problem characteristics, algorithmic, and systemic variations. Dynamic loop self-scheduling (DLS) techniques are devised to mitigate these factors, and consequently, improve application performance. On distributed-memory systems, DLS techniques can be implemented using a hierarchical master-worker exe…
Distributed Order Recording Techniques for Efficient Record-and-Replay of Multi-threaded Programs
Nanchang Hangkong University（南昌航空大学）；Pacific Northwest National Laboratory（太平洋西北国家实验室）；RIKEN R-CCS（理化学研究所计算科学研究中心）；NVIDIA（英伟达）；Lawrence Livermore National Laboratory（劳伦斯利弗莫尔国家实验室）；Technical University of Munich（慕尼黑工业大学）
Modern OpenMP threading techniques are used to convert the MPI-only Hartree-Fock code in the GAMESS program to a hybrid MPI/OpenMP algorithm. Two separate implementations that differ by the sharing or replication of key data structures among threads are considered, density and Fock matrices. All implementations are benchmarked on a super-computer of 3,000 Intel Xeon Phi processors. With 64 cores per processor, scaling numbers are reported on up to 192,000 cores. The hybrid MPI/OpenMP implementation reduces the memory footprint by approximately 200 times compared to the legacy code. The MPI/Ope…
New trends towards multiple core processors imply using standard programming models to develop efficient, reliable and portable programs for distributed memory multiprocessors and workstation PC clusters. Message passing using MPI is widely used to write efficient, reliable and portable applications. Control and data flow analysis concepts, techniques and tools are needed to understand and analyze MPI programs. If our point of interest is the program control and data flow analysis, to decide to parallelize or not to parallelize our applications, there is a question to be answered, " Can the ex…
MPI is the most widely used data transfer and communication model in High Performance Computing. The latest version of the standard, MPI-3, allows skilled programmers to exploit all hardware capabilities of the latest and future supercomputing systems. The revised asynchronous remote-memory-access model in combination with the shared-memory window extension, in particular, allow writing code that hides communication latencies and optimizes communication paths according to the locality of data origin and destination. The latter is particularly important for today's multi- and many-core systems.…
Upcoming HPC clusters will feature hybrid memories and storage devices per compute node. In this work, we propose to use the MPI one-sided communication model and MPI windows as unique interface for programming memory and storage. We describe the design and implementation of MPI storage windows, and present its benefits for out-of-core execution, parallel I/O and fault-tolerance. In addition, we explore the integration of heterogeneous window allocations, where memory and storage share a unified virtual address space. When performing large, irregular memory operations, we verify that MPI windo…
Most distributed-memory bulk-synchronous parallel programs in HPC assume that compute resources are available continuously and homogeneously across the allocated set of compute nodes. However, long one-off delays on individual processes can cause global disturbances, so-called idle waves, by rippling through the system. This process is mainly governed by the communication topology of the underlying parallel code. This paper makes significant contributions to the understanding of idle wave dynamics. We study the propagation mechanisms of idle waves across the ranks of MPI-parallel programs. We …
Python has become a dominant programming language for emerging areas like Machine Learning (ML), Deep Learning (DL), and Data Science (DS). An attractive feature of Python is that it provides easy-to-use programming interface while allowing library developers to enhance performance of their applications by harnessing the computing power offered by High Performance Computing (HPC) platforms. Efficient communication is key to scaling applications on parallel systems, which is typically enabled by the Message Passing Interface (MPI) standard and compliant libraries on HPC hardware. mpi4py is a Py…
