Moreno Marzolla

The gLite workload management system

The gLite Workload Management System (WMS) is a collection of components that provide the service responsible for distributing and managing tasks across computing and storage resources available on a Grid. The WMS basically receives requests of job execution from a client, finds the required appropriate resources, then dispatches and follows the jobs until completion, handling failure whenever possible. Other than single batch-like jobs, compound job types handled by the WMS are Directed Acyclic Graphs (a set of jobs where the input/output/execution of one of more jobs may depend on one or more other jobs), Parametric Jobs (multiple jobs with one parametrized description), and Collections (multiple jobs with a common description). Jobs are described via a flexible, high-level Job Definition Language (JDL). New functionality was recently added to the system (use of Service Discovery for obtaining new service endpoints to be contacted, automatic sandbox files archival/compression and sharing, support for bulk-submission and bulk-matchmaking). Intensive testing and troubleshooting allowed to dramatically increase both job submission rate and service stability. Future developments of the gLite WMS will be focused on reducing external software dependency, improving portability, robustness and usability.

Server consolidation in Clouds through gossiping

The success of Cloud computing, where computing power is treated as a utility, has resulted in the creation of many large datacenters that are very expensive to build and operate. In particular, the energy bill accounts for a significant fraction of the total operation costs. For this reason a significant attention is being devoted to energy conservation techniques, for example by taking advantage of the built-in power saving features of modern hardware. Cloud computing offers novel opportunities for achieving energy savings: Cloud systems rely on virtualization techniques to allocate computing resources on demand, and modern Virtual Machine (VM) monitors allow live migration of running VMs. Thus, energy conservation can be achieved through server consolidation, moving VM instances away from lightly loaded computing nodes so that they become empty and can be switched to low-power mode. In this paper we present V-MAN, a fully decentralized algorithm for consolidating VMs in large Cloud datacenters. V-MAN can operate on any arbitrary initial allocation of VMs on the Cloud, iteratively producing new allocations that quickly converge towards the one maximizing the number of idle hosts. V-MAN uses a simple gossip protocol to achieve efficiency, scalability and robustness to failures. Simulation experiments indicate that, starting from a random allocation, V-MAN produces an almost-optimal VM placement in just a few rounds; the protocol is intrinsically robust and can cope with computing nodes being added to or removed from the Cloud.

Performance evaluation of UML software architectures with multiclass Queueing Network models

Software performance based on performance models can be applied at early phases of the software development cycle to characterize the quantitative behavior of software systems. We propose an approach based on queueing networks models for performance prediction of software systems at the software architecture level, specified by UML. Starting from annotated UML Use Case, Activity and Deployment diagrams we derive a performance models based on multichain and multiclass Queueing Networks (QN). The UML model is annotated according to the UML Profile for Schedulability, Performance and Time Specification. The proposed algorithm translates the annotated UML specification into QN performance models, which can then be analyzed using standard solution techniques. Performance results are reported back at the software architecture level in the UML diagrams. As our approach can be fully automated and uses standard UML annotations, it can be integrated with other performance modeling approaches. Specifically, we discuss how this QN-based approach can be integrated with an existing simulation-based performance modeling tool.

Design and implementation of a P2P Cloud system

Cloud Computing has gained popularity in both research and industrial communities. Cloud users can acquire computing resources on a need basis, achieving on demand scalability; Cloud providers can maximize resource utilizations of datacenters, increasing their return on investments. While Cloud systems are usually hosted in large datacenters and are centrally managed, other types of Cloud architectures can be imagined. In this paper we describe the design and prototype implementation of a fully decentralized, P2P Cloud. A P2P Cloud allows organizations or even individual to build a computing infrastructure out of existing resources, which can be easily allocated among different tasks. We focus on the problem of maintaining a coherent structure over a set of unreliable computing resources. We show that gossip-based protocols can be used to maintain an overlay network on top of the computing nodes, and to partition the set of resources into multiple slices in such a way that the failure of individual nodes do not compromise the overall structure. Resource partitioning is one of the most important features of a Cloud, and therefore must be supported efficiently and reliably on any Cloud architecture. We describe a prototype Java implementation that is being developed to demonstrate the effectiveness of the proposed approach.

Spreading Processes in Multilayer Networks

Several systems can be modeled as sets of interconnected networks or networks with multiple types of connections, here generally called multilayer networks. Spreading processes such as information propagation among users of online social networks, or the diffusion of pathogens among individuals through their contact network, are fundamental phenomena occurring in these networks. However, while information diffusion in single networks has received considerable attention from various disciplines for over a decade, spreading processes in multilayer networks is still a young research area presenting many challenging research issues. In this paper, we review the main models, results and applications of multilayer spreading processes and discuss some promising research directions.

Performance prediction of web service workflows

Web Services play an important role in the Service-oriented Architecture paradigm, as they allow services to be selected on-the-fly to build applications out of existing components. In this scenario, the Business Process Execution Language notation can be used as an orchestration language which allows the user to describe interactions with Web Services in a standard way. The performance of a BPEL workflow is a very important factor for deciding which components must be selected, or to choose whether a given sequence of interactions can provide the requested quality of service. Due to its very dynamic nature, workflow performance evaluation can not be accomplished using traditional, heavy-weight techniques. In this paper we present a multi-view approach for the performance prediction of service-based applications encompassing both users and service provider(s) perspectives. As a first step towards the realization of this integrated framework we present an efficient approach for performance assessment of Web Service workflows described using the BPEL notation. Starting from annotated BPEL and WSDL specifications, we derive performance bounds on response time and throughput. In such a way users are able to assess the efficiency of a BPEL workflow, while service provider(s) can perform sizing studies or estimate performance gains of alternative upgrades to existing systems. To bring this approach to fruition we developed a prototype tool called bpel2qnbound, using which we analyze a simple case study.

A simulation-based approach to software performance modeling

Quantitative performance analysis of software systems should be integrated in the early stages of the development process. We propose a simulation-based performance modeling of software architectures specified in UML. We propose an algorithm for deriving a simulation model from annotated UML software architectures. We introduce the annotation for some UML diagrams, i.e., Use Case, Activity and Deployment diagrams, to describe system performance parameters. Then we show how to derive a process-oriented simulation model by automatically extracting information from the UML diagrams. Simulation provides performance results that are reported into the UML diagrams as tagged values. The proposed methodology has been implemented into a prototype tool called UML-?. The proposed methodology will be illustrated on a simple case study.

UML-PSI: the UML performance simulator

In this paper we describe UML-/spl Psi/, a software performance evaluation tool based on process-oriented simulation. The tool can be used to evaluate performances of software systems described as annotated UML diagrams. UML-/spl Psi/ transforms the software model into a performance model based on process-oriented simulation, executes the performance model and collects results. Performance results are inserted into the software model as tagged values associated to the relevant UML elements.

Resource Discovery in a Dynamic Grid Environment

Resource discovery in a grid environment is a critical problem, as a typical grid system includes a very large number of resources, which must be readily identified and accessed to run applications. Traditional grid discovery algorithms perform poorly, as they do not scale, nor allow grid-enabled applications to transparently query the whole set of grid resources. Peer-to-peer (P2P) has been argued as a suitable distributed paradigm that not only overcomes the issues of scalability of such discovery systems, but also better supports the discovery of resources in a context of dynamicity of resources and associated information. In this paper we propose a P2P system for indexing and discovering grid resources. We assume that grid resources are characterized by a set of attributes, and our system can be queried for resources satisfying arbitrary range conditions on these attributes. Note that traditional P2P searching techniques can not be directly applied in this case, since they work well mostly for static content and exact queries. Simulation results show that the system provides an adequate degree of scalability

Peer-to-peer systems for discovering resources in a dynamic grid

The convergence of the Grid and Peer-to-Peer (P2P) worlds has led to many solutions that try to efficiently solve the problem of resource discovery on Grids. Some of these solutions are extensions of P2P DHT-based networks. We believe that these systems are not flexible enough when the indexed data are very dynamic, i.e., the values of the resource attributes change very frequently over time. This is a common case for Grid metadata, like CPU loads, queue occupation, etc. Moreover, since common requests for Grid resources may be expressed as multi-attribute range queries, we think that the DHT-based P2P solutions are poorly flexible and efficient in handling them. In this paper we present two P2P systems. Both are based on Routing Indexes, which are used to efficiently route queries and update messages in the presence of highly variable data. The first system uses a tree-shaped overlay network. The second one is an evolution of the first, and is based on a two-level hierarchical network topology, where tree topologies must only be maintained at the lower level of the hierarchy, i.e., within the various node groups making up the network. The main goal of the second organization is to achieve a simpler maintenance of the overall P2P graph topology, by preserving the good properties of the tree-shaped topology. We discuss the results of extensive simulation studies aimed at assessing the performance and scalability of the proposed approaches. We also analyze how the network topologies affect the propagation of query and update messages.