Liria Matsumoto Sato

Exploiting idle cycles to execute data mining applications on clusters of PCs

In this paper we present and evaluate Inhambu, a distributed object-oriented system that supports the execution of data mining applications on clusters of PCs and workstations. This system provides a resource management layer, built on the top of Java/RMI, that supports the execution of the data mining tool called Weka. We evaluate the performance of Inhambu by means of several experiments in homogeneous, heterogeneous and non-dedicated clusters. The obtained results are compared with those achieved by a similar system named Weka-Parallel. Inhambu outperforms its counterpart for coarse grain applications, mainly for heterogeneous and non-dedicated clusters. Also, our system provides additional advantages such as application checkpointing, support for dynamic aggregation of hosts to the cluster, automatic restarting of failed tasks, and a more effective usage of the cluster. Therefore, Inhambu is a promising tool for efficiently executing real-world data mining applications. The software is delivered at the projects web site available at http://incubadora.fapesp.br/projects/inhambu/.

HyMPI – a MPI implementation for heterogeneous high performance systems

This paper presents the HyMPI, a runtime system to integrate several MPI implementations, used to develop Heterogeneous High Performance Applications. This means that a single image system can be composed by mono and multiprocessor nodes running several Operating Systems and MPI implementations, as well as, heterogeneous clusters as nodes of the system. HyMPI supports blocking and non-blocking point-to-point communication and collective communication primitive in order to increase the range of High Performance Applications that can use it and to keep compatibility with MPI Standard.

Performance Prediction Methodology for Parallel Programs with MPI in NOW Environments

We present a methodology for parallel programming, along with MPI performance measurement and prediction in a class of a distributed computing environments, namely networks of workstations. Our approach is based on a two-level model where, at the top, a new parallel version of timing graph representation is used to make explicit the parallel communication and code segments of a given parallel program, while at the bottom level, analytical models are developed to represent execution behavior of parallel communications and code segments. Execution time results obtained from execution, together with problem size and number of nodes, are input to the model, which allows us to predict the performance of similar cluster computing systems with a different number of nodes. The analytical model is validated by performing experiments over a homogeneous cluster of workstations. Final results show that our approach produces accurate predictions, within 5% of actual results.

On construction of a visualization toolkit for MPI parallel programs in cluster environments

The low cost and wide availability of PC-based clusters have made them an excellent alternative to access supercomputing. However, while network of workstations may be readily available, there is an increasing need for performance tools that support these platforms, in order to achieve even higher performance. One of possible ways to increase performance is parallel program restructuring. It is introduced in this paper a toolkit to generate graphical charts for visualization of MPI parallel programs, reflecting to its execution over time, with the use of DP*Graph representation, parallel version of timing graph. In other words, parallel programs are shown through charts its sequential codes, dependencies and communication structures in a particular cluster system platform. Still in this paper, it is discussed the implementation of this toolkit and present some experimental results obtained.

BSP cost and scalability analysis for MapReduce operations

Data abundance poses the need for powerful and easy‐to‐use tools that support processing large amounts of data. MapReduce has been increasingly adopted for over a decade by many companies, and more recently, it has attracted the attention of an increasing number of researchers in several areas. One main advantage is that the complex details of parallel processing, such as complex network programming, task scheduling, data placement, and fault tolerance, are hidden in a conceptually simple framework. MapReduce is supported by mature software technologies for deployment in data centers such as Hadoop. As MapReduce becomes popular for high‐performance applications, many questions arise concerning its performance and efficiency. In this paper, we demonstrated formally lower bounds on the isoefficiency function for MapReduce applications, when these applications can be modeled as BSP jobs. We also demonstrate how communication and synchronization costs can be dominant for MapReduce computations and discuss the conditions under which such scalability limits are valid. To our knowledge, this is the first study that demonstrates scalability bounds for MapReduce applications. We also discuss how some MapReduce implementations such as Hadoop can mitigate such costs to approach linear, or near‐to‐linear speedups. Copyright

An Architecture for Integrating Databases with Replication Support Based on the OGSA-DAI Middleware

In recent years, Grid Computing has become a key technologyin resource integration and sharing, thus improvingscientific collaboration among different institutions. Whilesome applications need access databases spread across theparticipants of the grid, OGSA-DAI is middleware that enablesdatabase integration and sharing from different vendors.By using replication techniques, this work aims toprovide a higher level of availability and reliability whenusing databases exposed to the grid. It defines an architecturefor systems that want to offer replication to the grid andwe implemented extensions to the OGSA-DAI middlewarein order to achieve this objective and prove the concepts ofthe proposed architecture. This paper presents the resultsobtained in the implementation and an initial study of itsapplication in a Health Care Information System.

Improvement on Scheduling Dependent Tasks for Grid Applications

Grid technology has been intensively developed lasting recent years, due to a need for high performance computing to process large amounts of data distributed around global networks. In addition to the improved performance obtained by parallel processing, this technology enables resources distributed among different nodes to be shared.In this study, the problem of transferring files that are generated during the execution of DAG workflows with interdependent tasks is addressed. The ineffectiveness of advanced file-transfer techniques in these cases is discussed, and a heuristic is proposed for dealing with the scheduling of interdependent and independent tasks that arrive on-line to be processes by grid infrastructure.A simulator was developed in order to simulate resources and tasks in a grid for testing the effectiveness of the new heuristic technique. The results are compared with those obtained with OLB showing that the proposed heuristic technique also leads to significant improvements in the makespan.

A novel strategy for building interoperable MPI environment in heterogeneous high performance systems

Breakthrough advances in microprocessor technology and efficient power management have altered the course of development of processors with the emergence of multi-core processor technology, in order to bring higher level of processing. The utilization of many-core technology has boosted computing power provided by cluster of workstations or SMPs, providing large computational power at an affordable cost using solely commodity components. Different implementations of message-passing libraries and system softwares (including Operating Systems) are installed in such cluster and multi-cluster computing systems. In order to guarantee correct execution of message-passing parallel applications in a computing environment other than that originally the parallel application was developed, review of the application code is needed. In this paper, a hybrid communication interfacing strategy is proposed, to execute a parallel application in a group of computing nodes belonging to different clusters or multi-clusters (computing systems may be running different operating systems and MPI implementations), interconnected with public or private IP addresses, and responding interchangeably to user execution requests. Experimental results demonstrate the feasibility of this proposed strategy and its effectiveness, through the execution of benchmarking parallel applications.

A message forward tool for integration of clusters of clusters based on MPI architecture

Advances in microprocessor technology, power management and network communication have altered the course of development of multiprocessor architectures in order to bring higher level of processing. The introduction of multi-core technology has boosted computing power provided by high-speed network of workstations and SMPs, providing large computational power at an affordable cost using solely commodity components. In this paper, it is presented a tool for integration of several clusters in a single High-Performance System based on MPI standard. The Gateway Process is responsible for MPI process communication channels control and message forwarding, through the use of a protocol that guarantees message ordering and sender/receiver synchronization. It is implemented to support system scalability, offering resources for point to point and collective operations. Results of experimental tests show that the proposed tool is practical and efficient.

Reducing Virtual Machine Live Migration Overhead via Workload Analysis

A key feature in virtualization technology is the Live Migration, which allows a Virtual Machine (VM) to be moved from a physical host to another without execution interruption. This feature enables the implementation of more sophisticated policies inside a cloud environment, such as energy and computational resources optimization, and improvement of quality-of-service. However live migration can impose severe performance degradation for the VM application and cause multiple impacts in service provider infrastructure, such as network congestion and co-located VM performance degradation. Different of several studies we consider the VM workload an important factor and we argue that carefully choosing a proper moment to migrate a VM can reduce the live migration penalties. This paper introduces a method to identify the workload cycles of a VM and based on that information it can postpone or, in some situations. In our experiments, using relevant benchmarks the proposed method was able to reduce up to 43% of network data transfer and reduce up to 74% of live migration time when compared to traditional consolidation strategies that perform live migration without considering the VM workload.