Predictive and distributed routing balancing (PR-DRB): high speed interconnection networks

Current parallel applications running on clusters require the use of an interconnection network to perform communications among all computing nodes available. Imbalance of communications can produce network congestion, reducing throughput and increasing latency, degrading the overall system performance. On the other hand, parallel applications running on these networks posses representative stages which allow their characterization, as well as repetitive behavior that can be identified on the basis of this characterization. This work presents the Predictive and Distributed Routing Balancing (PR-DRB), a new method developed to gradually control network congestion, based on paths expansion, traffic distribution and effective traffic load, in order to maintain low latency values. PR-DRB monitors messages latencies on intermediate routers, makes decisions about alternative paths and record communication pattern information encountered during congestion situation. Based on the concept of applications repetitiveness, best solution recorded are reapplied when saved communication pattern re-appears. Traffic congestion experiments were conducted in order to evaluate the performance of the method, and improvements were observed.

Properties and Algorithms of the KCube Interconnection Networks

The KCube interconnection network was first introduced in 2010 in order to exploit the good characteristics of two well-known interconnection networks, the hypercube and the Kautz graph. KCube links up multiple processors in a communication network with high density for a fixed degree. Since the KCube network is newly proposed, much study is required to demonstrate its potential properties and algorithms that can be designed to solve parallel computation problems. In this thesis we introduce a new methodology to construct the KCube graph. Also, with regard to this new approach, we will prove its Hamiltonicity in the general KC(m; k). Moreover, we will find its connectivity followed by an optimal broadcasting scheme in which a source node containing a message is to communicate it with all other processors. In addition to KCube networks, we have studied a version of the routing problem in the traditional hypercube, investigating this problem: whether there exists a shortest path in a Qn between two nodes 0n and 1n, when the network is experiencing failed components. We first conditionally discuss this problem when there is a constraint on the number of faulty nodes, and subsequently introduce an algorithm to tackle the problem without restrictions on the number of nodes.

Design and Analysis of Optical Interconnection Networks for a Dataflow Parallel Computer

This work shows the design, simulation, and analysis of two optical interconnection networks for a Dataflow parallel computer architecture. To verify the optical interconnection network performance on the Dataflow architecture, we have analyzed the load balancing among the processors during the parallel programs executions. The load balancing is a very important parameter because it is directly associated to the dataflow parallelism degree. This article proves that optical interconnection networks designed with simple optical devices can provide efficiently the dataflow requirements of a high performance communication system.

An oblivious shortest-path routing algorithm for fully connected cubic networks

Fully connected cubic networks (FCCNs) are a class of newly proposed hierarchical interconnection networks for multicomputer systems, which enjoy the strengths of constant node degree and good expandability. The shortest path routing in FCCNs is an open problem. In this paper, we present an oblivious routing algorithm for n-level FCCN with N = 8(n) nodes, and prove that this algorithm creates a shortest path from the source to the destination. At the costs of both an O(N)-parallel-step off-line preprocessing phase and a list of size N stored at each node, the proposed algorithm is carried out at each related node in O(n) time. In some cases the proposed algorithm is superior to the one proposed by Chang and Wang in terms of the length of the routing path. This justifies the utility of our routing strategy. (C) 2006 Elsevier Inc. All rights reserved.

Fault tolerant decentralised K-Means clustering for asynchronous large-scale networks

The K-Means algorithm for cluster analysis is one of the most influential and popular data mining methods. Its straightforward parallel formulation is well suited for distributed memory systems with reliable interconnection networks, such as massively parallel processors and clusters of workstations. However, in large-scale geographically distributed systems the straightforward parallel algorithm can be rendered useless by a single communication failure or high latency in communication paths. The lack of scalable and fault tolerant global communication and synchronisation methods in large-scale systems has hindered the adoption of the K-Means algorithm for applications in large networked systems such as wireless sensor networks, peer-to-peer systems and mobile ad hoc networks. This work proposes a fully distributed K-Means algorithm (EpidemicK-Means) which does not require global communication and is intrinsically fault tolerant. The proposed distributed K-Means algorithm provides a clustering solution which can approximate the solution of an ideal centralised algorithm over the aggregated data as closely as desired. A comparative performance analysis is carried out against the state of the art sampling methods and shows that the proposed method overcomes the limitations of the sampling-based approaches for skewed clusters distributions. The experimental analysis confirms that the proposed algorithm is very accurate and fault tolerant under unreliable network conditions (message loss and node failures) and is suitable for asynchronous networks of very large and extreme scale.

Reliable and Variation-Tolerant Interconnection Network for Low Power MPSOCS

Multi-Processor SoC (MPSOC) design brings to the foreground a large number of challenges, one of the most prominent of which is the design of the chip interconnection. With a number of on-chip blocks presently ranging in the tens, and quickly approaching the hundreds, the novel issue of how to best provide on-chip communication resources is clearly felt. Scaling down of process technologies has increased process and dynamic variations as well as transistor wearout. Because of this, delay variations increase and impact the performance of the MPSoCs. The interconnect architecture inMPSoCs becomes a single point of failure as it connects all other components of the system together. A faulty processing element may be shut down entirely, but the interconnect architecture must be able to tolerate partial failure and variations and operate with performance, power or latency overhead. This dissertation focuses on techniques at different levels of abstraction to face with the reliability and variability issues in on-chip interconnection networks. By showing the test results of a GALS NoC testchip this dissertation motivates the need for techniques to detect and work around manufacturing faults and process variations in MPSoCs’ interconnection infrastructure. As a physical design technique, we propose the bundle routing framework as an effective way to route the Network on Chips’ global links. For architecture-level design, two cases are addressed: (I) Intra-cluster communication where we propose a low-latency interconnect with variability robustness (ii) Inter-cluster communication where an online functional testing with a reliable NoC configuration are proposed. We also propose dualVdd as an orthogonal way of compensating variability at the post-fabrication stage. This is an alternative strategy with respect to the design techniques, since it enforces the compensation at post silicon stage.

Políticas de encaminamiento tolerantes a fallos

El uso intensivo y prolongado de computadores de altas prestaciones para ejecutar aplicaciones computacionalmente intensivas, sumado al elevado número de elementos que los componen, incrementan drásticamente la probabilidad de ocurrencia de fallos durante su funcionamiento. El objetivo del trabajo es resolver el problema de tolerancia a fallos para redes de interconexión de altas prestaciones, partiendo del diseño de polí­ticas de encaminamiento tolerantes a fallos. Buscamos resolver una determinada cantidad de fallos de enlaces y nodos, considerando sus factores de impacto y probabilidad de aparición. Para ello aprovechamos la redundancia de caminos de comunicación existentes, partiendo desde enfoques de encaminamiento adaptativos capaces de cumplir con las cuatro fases de la tolerancia a fallos: detección del error, contención del daño, recuperación del error, y tratamiento del fallo y continuidad del servicio. La experimentación muestra una degradación de prestaciones menor al 5%. En el futuro, se tratará la pérdida de información en tránsito.

Balanceo distribuido del encaminamiento para topologías fat-tree sobre redes Infiniband

Las redes de interconexión juegan un papel importante en el rendimiento de los sistemas de altas prestaciones. Actualmente la gestión del encaminamiento de los mensajes es un factor determinante para mantener las prestaciones de la red. Nuestra propuesta es trabajar sobre un algoritmo de encaminamiento adaptativo, que distribuye el encaminamiento de los mensajes para evitar los problemas de congestión en las redes de interconexión, que aparecen por el gran volumen de comunicaciones de aplicaciones científicas ó comerciales. El objetivo es ajustar el algoritmo a una topología muy utilizada en los sistemas actuales como lo es el fat‐tree, e implementarlo en una tecnología Infiniband. En la experimentación realizada comparamos el método de control de congestión de la arquitectura Infiniband, con nuestro algoritmo. Los resultados obtenidos muestran que mejoramos los niveles de latencia por encima de un 50% y de throughput entre un 38% y un 81%.

Modelo de red de interconexión basado en enlace

La computación de altas prestaciones es una área de la informática que evoluciona rápidamente, en la que actualmente aparecen nuevos computadores que llegan a los petaflops. Al principio del trabajo, se estudian los distintos tipos de redes de interconexión y los modelos de red que se utilizan para medir su latencia. El objetivo de este trabajo, es el diseño, implementación y simulación de un modelo de red de interconexión basado en enlace, que tiene en cuenta la información de topología y enrutamiento de la red de interconexión. Teniendo en cuenta que los modelos son una abstracción del sistema, en éste trabajo se hace la verificación y validación del modelo, para asegurar que éste se aproxima a lo planteado en el diseño y también que se parece al sistema que se quiere modelar.

Properties and Algorithms of the KCube Graphs

The KCube interconnection topology was rst introduced in 2010. The KCube graph is a compound graph of a Kautz digraph and hypercubes. Compared with the at- tractive Kautz digraph and well known hypercube graph, the KCube graph could accommodate as many nodes as possible for a given indegree (and outdegree) and the diameter of interconnection networks. However, there are few algorithms designed for the KCube graph. In this thesis, we will concentrate on nding graph theoretical properties of the KCube graph and designing parallel algorithms that run on this network. We will explore several topological properties, such as bipartiteness, Hamiltonianicity, and symmetry property. These properties for the KCube graph are very useful to develop efficient algorithms on this network. We will then study the KCube network from the algorithmic point of view, and will give an improved routing algorithm. In addition, we will present two optimal broadcasting algorithms. They are fundamental algorithms to many applications. A literature review of the state of the art network designs in relation to the KCube network as well as some open problems in this field will also be given.

‘Studies on some topics in product graphs’

For routing problems in interconnection networks it is important to find the shortest containers between any two vertices, since the w-wide diameter gives the maximum communication delay when there are up to w−1 faulty nodes in a network modeled by a graph. The concept of ‘wide diameter’ was introduced by Hsu [41] to unify the concepts of diameter and The concept of ‘domination’ has attracted interest due to its wide applications in many real world situations [38]. A connected dominating set serves as a virtual backbone of a network and it is a set of vertices that helps in routing. In this thesis, we make an earnest attempt to study some of these notions in graph products. This include, the diameter variability, the diameter vulnerability, the component factors and the domination criticality.connectivity

On the maximal connected component of hypercube with faulty vertices

In evaluating an interconnection network, it is indispensable to estimate the size of the maximal connected components of the underlying graph when the network begins to lose processors. Hypercube is one of the most popular interconnection networks. This article addresses the maximal connected components of an n -dimensional cube with faulty processors. We first prove that an n -cube with a set F of at most 2n - 3 failing processors has a component of size greater than or equal to2(n) - \F\ - 1. We then prove that an n -cube with a set F of at most 3n - 6 missing processors has a component of size greater than or equal to2(n) - \F\ - 2.

Generalized honeycomb torus is Hamiltonian

Generalized honeycomb torus is a candidate for interconnection network architectures, which includes honeycomb torus, honeycomb rectangular torus, and honeycomb parallelogramic torus as special cases. Existence of Hamiltonian cycle is a basic requirement for interconnection networks since it helps map a "token ring" parallel algorithm onto the associated network in an efficient way. Cho and Hsu [Inform. Process. Lett. 86 (4) (2003) 185-190] speculated that every generalized honeycomb torus is Hamiltonian. In this paper, we have proved this conjecture. (C) 2004 Elsevier B.V. All rights reserved.

On the maximal connected component of hypercube with faulty vertices (II)

evaluating the fault tolerance of an interconnection network, it is essential to estimate the size of a maximal connected component of the network at the presence of faulty processors. Hypercube is one of the most popular interconnection networks. In this paper, we prove that for ngreater than or equal to6, an n-dimensional cube with a set F of at most (4n-10) failing processors has a component of size greater than or equal to2"-\F-3. This result demonstrates the superiority of hypercube in terms of the fault tolerance.

A comparison-based diagnosis algorithm tailored for crossed cube multiprocessor systems

Comparison-based diagnosis is an effective approach to system-level fault diagnosis. Under the Maeng-Malek comparison model (NM* model), Sengupta and Dahbura proposed an O(N-5) diagnosis algorithm for general diagnosable systems with N nodes. Thanks to lower diameter and better graph embedding capability as compared with a hypercube of the same size, the crossed cube has been a promising candidate for interconnection networks. In this paper, we propose a fault diagnosis algorithm tailored for crossed cube connected multicomputer systems under the MM* model. By introducing appropriate data structures, this algorithm runs in O(Nlog(2)(2) N) time, which is linear in the size of the input. As a result, this algorithm is significantly superior to the Sengupta-Dahbura's algorithm when applied to crossed cube systems. (C) 2004 Elsevier B.V. All rights reserved.