Fast Broadcast by Message Segmentation

Microprocessors and Microsystems, 2002

A general-purpose multicomputer network must be able to ef®ciently handle broadcast communication because it is required by many parallel applications. Previous studies have focused mainly on the design of ef®cient broadcast algorithms. As a result, there has been hardly any work that assesses the performance of existing networks to handle this kind of communication. In an effort to ®ll this gap, this paper examines the relative performance merits of the low-dimensional k-ary n-cube and the hypermesh in the presence of broadcast traf®c. The former network has been one of the widely popular networks in current practical multicomputers, while the latter has recently been proposed, and shown to exhibit attractive topological properties. The results reveal that owing to its hypergraph topology, the hypermesh represents a potential candidate as a high-performance network for future multicomputers as it provides better support for broadcast communication than the k-ary n-cube.

Networks, 1995

Broadcasting refers to the process of dissemination of a set of messages originating from one node to all other nodes in a communication network. We assume that, at any given time, a node can transmit a message along at most one incident link and simultaneously receive a message along at most one incident link. We first present an algorithm for determining the amount of time needed to broadcast k messages in an arbitrary tree. Second, we show that, for every n, There exists a graph with n nodes whose k-message broadcast time matches the trivial lower bound ⌈ log n⌉ + k − 1 by designing a broadcast scheme for complete graphs. We call those graphs minimal broadcast graphs. Finally, we construct an n node minimal broadcast graph with fewer than (⌈log n⌉ + 1)2⌈ log n⌉ −1 edges.

HAL (Le Centre pour la Communication Scientifique Directe), 2016

2002

Data broadcasting as a means of efficient data dissemination is a key technology facilitating ubiquitous computing. For this reason, broadcast scheduling algorithms have received a lot of attention. However, all existing algorithms make the core assumption that the data items to be broadcast are immediately available in the transmitter's queue, ignoring the key role that the disk subsystem and the cache management play in the overall broadcast system performance. With this paper we contribute a comprehensive system's perspective towards the development of high performance broadcast systems, taking into account how broadcast scheduling, disk scheduling, and cache management algorithms affect the overall performance. We contribute novel techniques that ensure an efficient interplay between broadcast, cache management, and disk scheduling. We study comprehensively the performance of the broadcast server, as it consists of the broadcast scheduling, the disk scheduling, the cache management algorithms, and the transmitter. Our results show that the contributed algorithms yield considerably higher performance. Furthermore, one of our algorithms is shown to enjoy considerably higher performance, under all values of the problem and system parameters. A key contribution is the result that broadcast scheduling algorithms have only a small effect on the overall system performance, which necessitates the definition of different focal points for efforts towards high performance data broadcasting.

IEEE Transactions on Computers, 2000

The paper addresses ways in which one can use "broadcast communication" in distributed algorithms and the relevant issues of design and complexity. We present an algorithm for merging k sorted lists of n/k elements using k processors and prove its worst case complexity to be 2n, regardless of the number of processors, while neglecting the cost arising from possible conflicts on the broadcast channel. We also show that this algorithm is optimal under single-channel broadcast communication. In a variation of the algorithm, we show that by using an extra local memory of 0(k) the number of broadcasts is reduced to n. When the algorithm is used for sorting n elements with k processors, where each processor sorts its own list first and then merging, it has a complexity of 0(n/k log(n/k) + n), and is thus asymptotically optimal for large n. We also discuss the cost incurred by the channel access scheme and prove that resolving conflicts whenever k processors are involved introduces a cost factor of at least log k.

Many scientific applications running on distributed memory systems consume a substantial fraction of their total execution time exchanging data between processes. Thus, improving the performance of communication routines can significantly speed up overall application execution. In this context, we investigate the usage of a high-speed compression algorithm for floating-point data in the transmission of long messages, in particular broadcasts. We incorporated this algorithm in the broadcast primitive of an MPI library and evaluate the performance on different types of messages for up to 4096 processing cores. The results show that compression can significantly accelerate broadcasts.

2004

Modern high performance applications require efficient and scalable collective communication operations. Currently, most collective operations are implemented based on point-to-point operations. In this paper, we propose to use hardware multicast in InfiniBand to design fast and scalable broadcast operations in MPI. InfiniBand supports multicast with Unreliable Datagram (UD) transport service. This makes it hard to be directly used by an upper layer such as MPI. To bridge the semantic gap between MPI Bcast and InfiniBand hardware multicast, we have designed and implemented a substrate on top of InfiniBand which provides functionalities such as reliability, in-order delivery and large message handling. By using a sliding-window based design, we improve MPI Bcast latency by removing most of the overhead in the substrate out of the communication critical path. By using optimizations such as a new co-root based scheme and lazy ACK, we can further balance and reduce the overhead. We have also addressed many detailed design issues such as buffer management, efficient handling of out-of-order and duplicate messages, timeout and retransmission, flow control and RDMA based ACK communication.

Recent Advances in Parallel Virtual Machine and …, 2004

Improving communication performance is an important issue in cluster systems. This paper investigates the possibility of accelerating group communication at the level of message passing libraries. A new algorithm for implementing the broadcast communication primitive will be introduced. It enhances the performance of fully-switched cluster systems by using message decomposition and asynchronous communication. The new algorithm shows the dynamism and the portability of the software solutions, while it has a constant asymptotic time complexity achieved only with hardware support before. Test measurements show that the algorithm really has a constant time complexity, and in certain cases it can outperform the widely used binary tree approach by 100 percent. The presented algorithm can be used to increase the performance of broadcasting, and can also indirectly speed up various group communication primitives used in standard message passing libraries.

2005 International Conference on Parallel Processing Workshops (ICPPW'05), 2005

Broadcast Communication is among the most primitive collective capabilities of any message passing network. Broadcast algorithms for the mesh have been widely reported in the literature. However, most existing algorithms have been studied within limited conditions, such as light traffic load and fixed network sizes. In other words, most of these algorithms have not been studied at different Quality of Service (QoS) levels. In contrast, this study examines the broadcast operation, taking into account the scalability, parallelism, a wide range of traffic loads through the propagation of broadcast messages. To the best of our knowledge, this study is the first to consider the issue of broadcast latency at both the network and node levels across different traffic loads. Results are shown from a comparative analysis confirming that the coded-path based broadcast algorithms exhibit superior performance characteristics over some existing algorithms.

Journal of Systems Architecture, 2005

Maximising the performance of parallel systems requires matching message-passing algorithms and application characteristics with a suitable underling interconnection network. Broadcast algorithms for wormhole-switched meshes have been widely reported in the literature. However, most of these algorithms handle broadcast in a sequential manner and do not scale well with the network size. As a consequence, many parallel applications cannot be efficiently supported using existing techniques. Motivated by these observations, this paper presents a new efficient broadcast algorithm for the mesh, called the Plane-Based (PB) algorithm. The main feature of this approach is its ability to perform broadcast operation with a high degree of scalability and parallelism. Furthermore, performance is insensitive to the network size, i.e., only three message-passing steps are required to implement a broadcast operation irrespective of the network size. Results from a comparative analysis demonstrate that the PB algorithm exhibits superior performance characteristics over those of the well-known Recursive Doubling and Extending Dominating Node algorithms.