A Study of Process Arrival Patterns for MPI Collective Operations

Process arrival pattern, which denotes the timing when different processes arrive at an MPI collective operation, can have a significant impact on the performance of the operation. In this work, we characterize the process arrival patterns in a set of MPI programs on two common cluster platforms, use a micro-benchmark to study the process arrival patterns in MPI programs with balanced loads, and investigate the impacts of different process arrival patterns on collective algorithms. Our results show that (1) the differences between the times when different processes arrive at a collective operation are usually sufficiently large to affect the performance; (2) application developers in general cannot effectively control the process arrival patterns in their MPI programs in the cluster environment: balancing loads at the application level does not balance the process arrival patterns; and (3) the performance of collective communication algorithms is sensitive to process arrival patterns. These results indicate that process arrival pattern is an important factor that must be taken into consideration in developing and optimizing MPI collective routines. We propose a scheme that achieves high performance with different process arrival patterns, and demonstrate that by explicitly considering process arrival pattern, more efficient MPI collective routines than the current ones can be obtained.     

High Performance RDMA-Based MPI Implementation over InfiniBand

Although InfiniBand Architecture is relatively new in the high performance computing area, it offers many features which help us to improve the performance of communication subsystems. One of these features is Remote Direct Memory Access (RDMA) operations. In this paper, we propose a new design of MPI over InfiniBand which brings the benefit of RDMA to not only large messages, but also small and control messages. We also achieve better scalability by exploiting application communication pattern and combining send/receive operations with RDMA operations. Our RDMA-based MPI implementation achieves a latency of 6.8 sec for small messages and a peak bandwidth of 871 million bytes/sec. Performance evaluation shows that for small messages, our RDMA-based design can reduce the latency by 24%, increase the bandwidth by over 104%, and reduce the host overhead by up to 22% compared with the original design. For large data transfers, we improve performance by reducing the time for transferring control messages. We have also shown that our new design is beneficial to MPI collective communication and NAS Parallel Benchmarks.     

A framework for adaptive collective communications for heterogeneous hierarchical computing systems

Collective communication operations are widely used in MPI applications and play an important role in their performance. However, the network heterogeneity inherent to grid environments represent a great challenge to develop efficient high performance computing applications. In this work we propose a generic framework based on communication models and adaptive techniques for dealing with collective communication patterns on grid platforms. Toward this goal, we address the hierarchical organization of the grid, selecting the most efficient communication algorithms at each network level. Our framework is also adaptive to grid load dynamics since it considers transient network characteristics for dividing the nodes into clusters. Our experiments with the broadcast operation on a real-grid setup indicate that an adaptive framework allows significant performance improvements on MPI collective communications.     

Application-oriented ping-pong benchmarking: how to assess the real communication overheads

Moving data between processes has often been discussed as one of the major bottlenecks in parallel computing—there is a large body of research, striving to improve communication latency and bandwidth on different networks, measured with ping-pong benchmarks of different message sizes. In practice, the data to be communicated generally originates from application data structures and needs to be serialized before communicating it over serial network channels. This serialization is often done by explicitly copying the data to communication buffers. The message passing interface (MPI) standard defines derived datatypes to allow zero-copy formulations of non-contiguous data access patterns. However, many applications still choose to implement manual pack/unpack loops, partly because they are more efficient than some MPI implementations. MPI implementers on the other hand do not have good benchmarks that represent important application access patterns. We demonstrate that the data serialization can consume up to 80 % of the total communication overhead for important applications. This indicates that most of the current research on optimizing serial network transfer times may be targeted at the smaller fraction of the communication overhead. To support the scientific community, we extracted the send/recv-buffer access patterns of a representative set of scientific applications to build a benchmark that includes serialization and communication of application data and thus reflects all communication overheads. This can be used like traditional ping-pong benchmarks to determine the holistic communication latency and bandwidth as observed by an application. It supports serialization loops in C and Fortran as well as MPI datatypes for representative application access patterns. Our benchmark, consisting of seven micro-applications, unveils significant performance discrepancies between the MPI datatype implementations of state of the art MPI implementations. Our micro-applications aim to provide a standard benchmark for MPI datatype implementations to guide optimizations similarly to the established benchmarks SPEC CPU and Livermore Loops.     

Performance analysis and optimization of MPI collective operations on multi-core clusters

Memory hierarchy on multi-core clusters has twofold characteristics: vertical memory hierarchy and horizontal memory hierarchy. This paper proposes new parallel computation model to unitedly abstract memory hierarchy on multi-core clusters in vertical and horizontal levels. Experimental results show that new model can predict communication costs for message passing on multi-core clusters more accurately than previous models, only incorporated vertical memory hierarchy. The new model provides the theoretical underpinning for the optimal design of MPI collective operations. Aimed at horizontal memory hierarchy, our methodology for optimizing collective operations on multi-core clusters focuses on hierarchical virtual topology and cache-aware intra-node communication, incorporated into existing collective algorithms in MPICH2. As a case study, multi-core aware broadcast algorithm has been implemented and evaluated. The results of performance evaluation show that the above methodology for optimizing collective operations on multi-core clusters is efficient.     

The scalable process topology interface of MPI 2.2

The Message‐passing Interface (MPI) standard provides basic means for adaptations of the mapping of MPI process ranks to processing elements to better match the communication characteristics of applications to the capabilities of the underlying systems. The MPI process topology mechanism enables the MPI implementation to rerank processes by creating a new communicator that reflects user‐supplied information about the application communication pattern. With the newly released MPI 2.2 version of the MPI standard, the process topology mechanism has been enhanced with new interfaces for scalable and informative user‐specification of communication patterns. Applications with relatively static communication patterns are encouraged to take advantage of the mechanism whenever convenient by specifying their communication pattern to the MPI library. Reference implementations of the new mechanism can be expected to be readily available (and come at essentially no cost), but non‐trivial implementations pose challenging problems for the MPI implementer. This paper is first and foremost addressed to application programmers wanting to use the new process topology interfaces. It explains the use and the motivation for the enhanced interfaces and the advantages gained even with a straightforward implementation. For the MPI implementer, the paper summarizes the main issues in the efficient implementation of the interface and explains the optimization problems that need to be (approximately) solved by a good MPI library. Copyright © 2010 John Wiley & Sons, Ltd.     

PC机群上JIAJIA与MPI的比较

对JIAJIA和MPI (message passing interface)是进行了比较.JIAJIA和MPI分别代表共享存储和消息传递的编程模式.MPI显式进行数据传输,编程复杂;JIAJIA由底层维护数据一致性,并附加提供简单的消息传递函数,编程容易、灵活.JIAJIA分配共享内存时开销较大,初始化时间比MPI长.提出了一个关于并行加速比与进程数目之间关系的近似经验公式,推出JIAJIA和MPI性能差距随着进程数目的增多而增大的结论.测试结果表明,大部分应用程序的JIAJIA和MPI版本的并行性能差距不超过10%.对于通信量很小的应用程序,其JIAJIA和MPI的性能差距较小,而通信量本身较大的应用程序,其JIAJIA和MPI的性能差距主要取决于运行时产生的实际通信量.     

一种高性能的全序组播算法

全序组播是构建分布式应用程序的一种重要组通信原语,它能够保证一个通信组中的所有成员都按照同样的顺序接收消息.目前的全序组播算法不能同时获取低延迟和高吞吐量,并且缺乏对应用程序通信模式的适应性,因此不适用于高性能计算环境.在分析已有算法排序机制基础上,指出影响全序组播算法性能的关键因素,并提出一种基于leader/followers模式和阻塞检测机制的新算法.算法工作原理如下:每一个组成员都可以在任意时刻发送消息,但只能提交来自当前leader成员的消息;一旦leader成员进入不活跃状态,则通过特殊的命令来指定某个活跃的follower成员为新的leader成员.模拟实验结果表明,该算法在延迟时间和吞吐量等性能指标方面都优于已有算法,同时在突发消息模式下能够大幅度提升性能.     

HARNESS fault tolerant MPI design, usage and performance issues

Initial versions of MPI were designed to work efficiently on multi-processors which had very little job control and thus static process models. Subsequently forcing them to support a dynamic process model suitable for use on clusters or distributed systems would have reduced their performance. As current HPC collaborative applications increase in size and distribution the potential levels of node and network failures increase. This is especially true when MPI implementations are used as the communication media for GRID applications where the GRID architectures themselves are inherently unreliable thus requiring new fault tolerant MPI systems to be developed. Here we present a new implementation of MPI called FT-MPI that allows the semantics and associated modes of failures to be explicitly controlled by an application via a modified MPI API. Given is an overview of the FT-MPI semantics, design, example applications and some performance issues such as efficient group communications and complex data handling. Also briefly described is the HARNESS g_hcore system that handles low-level system operations on behalf of the MPI implementation. This includes details of plug-in services developed and their interaction with the FT-MPI runtime library.     

Evaluating InfiniBand performance with PCI Express

The InfiniBand architecture is an industry standard that offers low latency and high bandwidth as well as advanced features such as remote direct memory access (RDMA), atomic operations, multicast, and quality of service. InfiniBand products can achieve a latency of several microseconds for small messages and a bandwidth of 700 to 900 Mbytes/s. As a result, it is becoming increasingly popular as a high-speed interconnect technology for building high-performance clusters. The Peripheral Component Interconnect (PCI) has been the standard local-I/O-bus technology for the last 10 years. However, more applications require lower latency and higher bandwidth than what a PCI bus can provide. As an extension, PCI-X offers higher peak performance and efficiency. InfiniBand host channel adapters (HCAs) with PCI Express achieve 20 to 30 percent lower latency for small messages compared with HCAs using 64-bit, 133-MHz PCI-X interfaces. PCI Express also improves performance at the MPI level, achieving a latency of 4.1/spl mu/s for small messages. It can also improve MPI collective communication and bandwidth-bound MPI application performance.     

Dynamic-CoMPI: dynamic optimization techniques for MPI parallel applications

This work presents an optimization of MPI communications, called Dynamic-CoMPI, which uses two techniques in order to reduce the impact of communications and non-contiguous I/O requests in parallel applications. These techniques are independent of the application and complementaries to each other. The first technique is an optimization of the Two-Phase collective I/O technique from ROMIO, called Locality aware strategy for Two-Phase I/O (LA-Two-Phase I/O). In order to increase the locality of the file accesses, LA-Two-Phase I/O employs the Linear Assignment Problem (LAP) for finding an optimal I/O data communication schedule. The main purpose of this technique is the reduction of the number of communications involved in the I/O collective operation. The second technique, called Adaptive-CoMPI, is based on run-time compression of MPI messages exchanged by applications. Both techniques can be applied on every application, because both of them are transparent for the users. Dynamic-CoMPI has been validated by using several MPI benchmarks and real HPC applications. The results show that, for many of the considered scenarios, important reductions in the execution time are achieved by reducing the size and the number of the messages. Additional benefits of our approach are the reduction of the total communication time and the network contention, thus enhancing, not only performance, but also scalability.     

非平衡进程到达模式下MPI广播的性能优化方法

为了提高非平衡进程到达(unbalanced process arrival,简称UPA)模式下MPI广播的性能,对UPA模式下的广播问题进行了理论分析,证明了在多核集群环境中通过节点内多个MPI进程的竞争可以有效减少UPA对MPI广播性能的影响,并在此基础上提出了一种新的优化方法,即竞争式流水化方法(competitive and pipelined method,简称CP).CP方法通过一种节点内进程竞争机制在广播过程中尽早启动节点间通信,经该方法优化的广播算法利用共享内存在节点内通信,利用由竞争机制产生的引导进程执行原算法在节点间通信.并且,该方法使节点间通信和节点内通信以流水方式重叠执行,能够有效利用集群系统各节点的多核优势,减少了MPI广播受UPA的影响,提高了性能.为了验证CP方法的有效性,基于此方法优化了3种典型的MPI广播算法,分别适用于不同消息长度的广播.在真实系统中,通过微基准测试和两个实际的应用程序对CP广播进行了性能评价,结果表明,该方法能够有效地提高传统广播算法在UPA模式下的性能.在应用程序的负载测试实验结果中,CP广播的性能较流水化广播的性能提高约16%,较MVAPICH21.2中广播的性能提高18%～24%.     

基于天河互连MPI聚合通信归约操作卸载优化

MPI collective communication operation is widely used in parallel scientific application, which has an important influence impact on the scalabilityof the program. Tianhe interconnect network supports the trigger communication operations, which can offload the messaging and processing work and improve the performance between nodes. Allreduce and Reduce algorithms under different tree topological structures are designed by using thetriggered operations to lower the latency the reduction operation communication between nodes. Tests based on the actual system platform show that that, compared with the point to point implementation of these two types of operations in MPICH, the offload algorithm based on trigger can reduce the running time by up to 59.6% at different node scales.     

Performance evaluation of a Windows NT based PC cluster for high performance computing

In recent times the computational power of personal computers has remarkably increased and the use of groups of PCs and workstations, connected by a network and dedicated to parallel computations, is today frequent. Computing clusters are mainly based on UNIX workstations and Linux PCs but, in the last few years, different implementations of message passing systems were made available also for Microsoft Windows. In this work we test the performance of two implementations of MPI for Windows platforms, and we compare the results with those obtained from Linux systems.     

Towards Scalable Java HPC with Hybrid and Native Communication Devices in MPJ Express

MPJ Express is a messaging system that allows application developers to parallelize their compute-intensive sequential Java codes on High Performance Computing clusters and multicore processors. In this paper, we extend MPJ Express software to provide two new communication devices. The first device—called hybrid—enables MPJ Express to exploit hybrid parallelism on cluster of multicore processors by sitting on top of existing shared memory and network communication devices. The second device—called native—uses JNI wrappers in interfacing MPJ Express to native MPI implementations like MPICH and Open MPI. We evaluate performance of these devices on a range of interconnects including 1G/10G Ethernet, 10G Myrinet and 40G InfiniBand. In addition, we analyze and evaluate the cost of MPJ Express buffering layer and compare it with the performance numbers of other Java MPI libraries. Our performance evaluation reveals that the native device allows MPJ Express to achieve comparable performance to native MPI libraries—for latency and bandwidth of point-to-point and collective communications—which is a significant gain in performance compared to existing communication devices. The hybrid communication device—without any modifications at application level—also helps parallel applications achieve better speedups and scalability by exploiting multicore architecture. Our performance evaluation quantifies the cost incurred by buffering and its impact on overall performance of software. We witnessed comparative performance as both new devices improve application performance and achieve upto 90 % of the theoretical bandwidth available without application rewriting effort—including NAS Parallel Benchmarks, point-to-point and collective communication.     

Fat-tree routing and node ordering providing contention free traffic for MPI global collectives

As the size of High Performance Computing clusters grows, so does the probability of interconnect hot spots that degrade the latency and effective bandwidth the network provides. This paper presents a solution to this scalability problem for real life constant bisectional-bandwidth fat-tree topologies. It is shown that maximal bandwidth and cut-through latency can be achieved for MPI global collective traffic. To form such a congestion-free configuration, MPI programs should utilize collective communication, MPI-node-order should be topology aware, and the packet routing should match the MPI communication patterns. First, we show that MPI collectives can be classified into unidirectional and bidirectional shifts. Using this property, we propose a scheme for congestion-free routing of the global collectives in fully and partially populated fat trees running a single job. The no-contention result is then obtained for multiple jobs running on the same fat-tree by applying some job size and placement restrictions. Simulation results of the proposed routing, MPI-node-order and communication patterns show no contention which provides a 40% throughput improvement over previously published results for all-to-all collectives.     

MPI Correctness Checking for OpenMP/MPI Applications

The MPI interface is the de-facto standard for message passing applications, but it is also complex and defines several usage patterns as erroneous. A current trend is the investigation of hybrid programming techniques that use MPI processes and multiple threads per process. As a result, more and more MPI implementations support multi-threading, which are restricted by several rules of the MPI standard. In order to support developers of hybrid MPI applications, we present extensions to the MPI correctness checking tool Marmot. Basic extensions make it aware of OpenMP multi-threading, while further ones add new correctness checks. As a result, it is possible to detect errors that actually occur in a run with Marmot. However, some errors only occur for certain execution orders, thus, we present a novel approach using artificial data races, which allows us to employ thread checking tools, e.g., Intel Thread Checker, to detect MPI usage errors.     

Performance Modeling and Evaluation of MPI

Users of parallel machines need to have a good grasp for how different communication patterns and styles affect the performance of message-passing applications. LogGP is a simple performance model that reflects the most important parameters required to estimate the communication performance of parallel computers. The message passing interface (MPI) standard provides new opportunities for developing high performance parallel and distributed applications. In this paper, we use LogGP as a conceptual framework for evaluating the performance of MPI communications on three platforms: Cray-Research T3D, Convex Exemplar 1600SP, and a network of workstations (NOW). We develop a simple set of communication benchmarks to extract the LogGP parameters. Our objective in this is to compare the performance of MPI communication on several platforms and to identify a performance model suitable for MPI performance characterization. In particular, two problems are addressed: how LogGP quantifies MPI performance and what extra features are required for modeling MPI, and how MPI performance compare on the three computing platforms: Cray Research T3D, Convex Exemplar 1600SP, and workstations clusters.     

Protocol Customization for Improving MPI Performance on RDMA-Enabled Clusters

Optimizing Message Passing Interface (MPI) point-to-point communication for large messages is of paramount importance since most communications in MPI applications are performed by such operations. Remote Direct Memory Access (RDMA) allows one-sided data transfer and provides great flexibility in the design of efficient communication protocols for large messages. However, achieving high point-to-point communication performance on RDMA-enabled clusters is challenging due to both the complexity in communication protocols and the impact of the protocol invocation scenario on the performance of a given protocol. In this work, we analyze existing protocols and show that they are not ideal in many situations, and propose to use protocol customization, that is, different protocols for different situations to improve MPI performance. More specifically, by leveraging the RDMA capability, we develop a set of protocols that can provide high performance for all protocol invocation scenarios. Armed with this set of protocols that can collectively achieve high performance in all situations, we demonstrate the potential of protocol customization by developing a trace-driven toolkit that allows the appropriate protocol to be selected for each communication in an MPI application to maximize performance. We evaluate the performance of the proposed techniques using micro-benchmarks and application benchmarks. The results indicate that protocol customization can out-perform traditional communication schemes by a large degree in many situations.     

Towards Efficient Execution of MPI Applications on the Grid: Porting and Optimization Issues

The message passing interface (MPI) is a standard used by many parallel scientific applications. It offers the advantage of a smoother migration path for porting applications from high performance computing systems to the Grid. In this paper Grid-enabled tools and libraries for developing MPI applications are presented. The first is MARMOT, a tool that checks the adherence of an application to the MPI standard. The second is PACX-MPI, an implementation of the MPI standard optimized for Grid environments. Besides the efficient development of the program, an optimal execution is of paramount importance for most scientific applications. We therefore discuss not only performance on the level of the MPI library, but also several application specific optimizations, e.g., for a sparse, parallel equation solver and an RNA folding code, like latency hiding, prefetching, caching and topology-aware algorithms.