Ana Ripoll

A distributed diffusion method for dynamic load balancing on parallel computers

Parallel applications can be divided into tasks that can be executed simultaneously in different processors. Depending on prior knowledge about computational requirements of the problem, the assignment of tasks to processors can be guided in two ways: static and dynamic. We propose a new dynamic load balancing algorithm based on the diffusion approach which employs overlapping balancing domains to achieve global balancing. Since current diffusion methods consider discrete units, the algorithms may produce solutions which, although they are locally balanced prove to be globally unbalanced. Our method solves this problem taking into account the load maximum difference between two processors within each domain, providing a more efficient load balancing process.<<ETX>>

A New Task Graph Model for Mapping Message Passing Applications

The exploitation of parallelism in a message passing platform implies a previous modeling phase of the parallel application as a task graph, which properly reflects its temporal behavior. In this paper, we analyze the classical task graph models of the literature and their drawbacks when modeling message passing programs with an arbitrary task structure. We define a new task graph model called temporal task interaction graph (TTIG) that integrates the classical models used in the literature. The TTIG allows us to explicitly capture the ability of concurrency of adjacent tasks for applications where adjacent tasks can communicate at any point inside them. A mapping strategy is developed from this model, which minimizes the expected execution time by properly exploiting task parallelism. The effectiveness of this approach has been proved in different experimentation scopes for a wide range of message passing applications.

An asynchronous and iterative load balancing algorithm for discrete load model

Diffusion algorithms are some of the most popular algorithms for dynamic load balancing in which loads move from heavily loaded processors to lightly loaded neighbor processors. To achieve a global load balance in a parallel computer, the algorithm is iterated until the load difference between any two processors is smaller than a specified value. Therefore, one fundamental property to be studied is algorithm convergence. Several analytical works on the convergence of different diffusion load balancing algorithms have been carried out, but they treat loads as non-negative real quantities. In this paper, we describe the Diffusion Algorithm Searching Unbalanced Domains (DASUD) algorithm, which uses loads as non-negative integer values and, unlike existing algorithms, reaches a local balance situation where the maximum load difference between any two processor in the set of neighbor processors for each processor is one load unit. The convergence property of an asynchronous implementation of DASUD using integer loads is proven theoretically.

Performance comparison of dynamic load-balancing strategies for distributed computing

The DASUD (Diffusion Algorithm Searching Unbalanced Domains) algorithm belongs to the nearest-neighbours class and operates in a diffusion scheme where a processor balances its load with all its neighbours. DASUD detects unbalanced domains and performs local exchange of load between processors to achieve global balancing. The DASUD algorithm has been evaluated by comparison with another well-known strategy, namely, the SID (Sender Initiated Diffusion) algorithm across a range of network topologies including ring, torus and hypercube where the number of processors varies from 8 to 128. From the experiments we have observed that DASUD outperforms the other strategy as it provides the best trade-of-between the balance degree obtained at the final state and the number of iterations required to reach such a state. DASUD is able to coerce any initial load distribution into a highly balanced global state and also exhibits good scalability properties.

The Convergence of Realistic Distributed Load-Balancing Algorithms

We give a general model of partially asynchronous, distributed load-balancing algorithms for the discrete load model in parallel computers, where the processor loads are treated as non-negative integers. We prove that all load-balancing algorithms in this model are finite. This means that all load-balancing algorithms based on this model are guaranteed to reach a stable situation at a certain time (which depends on the particular algorithm) at which no load will be sent from one processor to another. With an additional assumption, we prove that the largest load difference between any two processors, in the final stable situation of the load-balancing algorithms in this model, is upper-bounded by the diameter of the topology.

Designing a Video-on-Demand System for a Brazilian High Speed Network

Despite the availability of video-on-demand (VoD) services in a number of cities around the world, the large-scale deployment of VoD services has not been widely explored due to economic constraints. The high complexity founded in these systems makes the design process very difficult. This study presents a VoD system design for a highspeed metropolitan network, based on ATM technology, located in Brazil. To accomplish this goal, we have utilized our own simulation-based tool for VoD design called VODSim. This tool allows VoD research on any VoD platform, simulating the complete flow path followed by requests from clients to servers passing through the network infrastructure, allowing the configuration and tuning of multiple system parameters. Using VODSim, we analyse possible VoD architectures for the proposal environment (i.e. Centralized-server, Independent-servers, One-level-proxies and Distributedservers), studying key-parameters that have influence in system design. Experimental results show that Distributed-server architecture is the one that adjusts better to the analyzed infrastructure, improving streaming capacity by 16%, 25% and 28% more than One-level-proxies, Centralized and Independent-servers respectively. Taking in account future growth in VoD system, the Distributed-server even provides an unlimited scalability through new local network incorporation and using P2P techniques.

A new model for static mapping of parallel applications with task and data parallelism

The efficient mapping of parallel tasks is essential in order to exploit the gain from parallelisation. In this work, we focus on modelling and mapping message-passing applications that are defined by the programmer with an arbitrary interaction pattern among tasks. A new model is proposed, known as TTIG (Temporal Task Interaction Graph), which captures not only computation and communication costs, but also the percentages of concurrency between tasks. From this model, a mapping strategy is developed that minimises expected execution time by properly exploiting task parallelism. The effectiveness of this approach has been proven for a real image-processing application on a cluster of PCs.

Clustering and reassignment-based mapping strategy for message-passing architectures

A fundamental issue affecting the performance of a parallel program is the assignment of tasks to processors in order to get the minimum completion time. We present a compilation-time two-stage mapping strategy (denoted as CREMA) used for mapping arbitrary programs (modeled as TIG graphs) onto message passing parallel systems with any architecture. In contrast to most of the other approaches found in the literature, CREMA is not tied to any particular architecture or any specific algorithm. The first stage is based on task clustering and task reassignment algorithms that contract the original task graph. The second stage takes the contracted graph and tries to successfully match the physical properties of the target system. It has been evaluated for a wide range of both regular and irregular graphs that correspond to some well-known real applications. The results show that CREMA provides a good trade-off between mapping quality and computational complexity.

Distributed P2P merging policy to decentralize the multicasting delivery

Advances in network technology make multicast one of the most feasible video streaming delivery techniques for the near future. However, the scalability of a multicast VoD system is limited by the server bandwidth. In this paper, we propose a new multicast delivery scheme that allows every active client to collaborate with the server in order to scale the VoD system performance beyond the servers physical limitations. The solution combined the multicast delivery scheme and peer-2-peer paradigm in order to decentralize the delivery process. The new video delivery scheme is able to merge two or more multicast channels using distributed collaborations between a group of clients. We compared the new policy with chaining and patching schemes and the experimental results showed that our policy is better than previous schemes in terms of reduction of resource requirements and local network load. Compared with multicast patching policy, the new scheme reduced the resource requirement up to 77.5% while the local network load was 66.9% lower than a peer-2-peer chaining policy.

Modelling message-passing programs for static mapping

An efficient mapping of a parallel program in the processors is vital for achieving a high performance on a parallel computer. When the structure of the parallel program in terms of its task execution times, task dependencies, and amount communication data, is known a priori, mapping can be accomplished statically at compile time. Mapping algorithms start from a parallel application model and map automatically tasks to processors in order to minimise the execution time of the program. In this paper we discuss the current models used in mapping parallel programs: Task Precedence Graph (TPG), Task Interaction Graph (TIG) and we define a new model called Temporal Task Interaction Graph (TTIG). The contribution of the TTIG is that it enhances these two previous models with the ability to explicitly capture the potential degree of parallel execution between adjacent tasks allowing the development of efficient mapping algorithms. Experimentation had been performed in order to show the effectiveness of TTIG model for a set of graphs. The results are compared with the optimal assignment and the obtained using TIG model and they confirm that using the TTIG model, better assignments can be obtained.