Jonathan Dukes

Using Dynamic Replication to Manage Service Availability in a Multimedia Server Cluster

Dynamic replication policies assign non-disjoint subsets of multimedia presentations to nodes in a server cluster, replicating selected presentations to achieve load-balancing, while avoiding complete replication of the multimedia archive on every node. This paper presents a development of our existing Dynamic RePacking policy, which creates a configurable minimum number of replicas of selected presentations, increasing their availability. These additional replicas are assigned to nodes in a manner that allows load-balancing to be maintained when nodes fail. By separating replication to achieve load-balancing from replication to increase the availability of individual presentations, the trade-off between availability and storage cost can be controlled. This is illustrated by performance results from a prototype multimedia server cluster, which uses a group-communication service to implement inter-node communication.

Supporting job-level secure access to GPGPU resources on existing grid infrastructures

Grids provide secure, utility-like access to a wide variety of large-scale, distributed computational and storage resources. In particular, the European Grid Infrastructure (EGI) and Open Science Grid (OSG) have excelled in processing vast workloads of independent jobs for the research community. Researchers demand increasingly faster processing speeds to solve increasingly larger and more complex problems. To meet this need, attention has shifted over the past decade away from single-core processing models towards the use of multi-core, many-core and massively parallel computational accelerators. The increasing availability and use of General Purpose Graphic Processing Units (GPGPUs) are an example of this. This paper addresses many of the challenges that exist in the integration of resources such as GPGPUs into Grid infrastructures. Specifically, solutions are proposed for discovering and describing GPGPU Grid resources, specifying multi-GPGPU job requirements, performing multi-GPGPU allocation to jobs, dynamically updating publicly-readable GPGPU usage information and enforcing GPGPU access control to prevent distinct jobs from inadvertently accessing the same device. The proposed solution is fully compatible with widely-used and accepted standards and middleware including the GLUE 2.0 schema and EGI Unified Middleware Distribution. A prototype implementation is also described.

Re-evaluating Multicast Streaming Using Large-Scale Network Simulation

Internet-based media streaming services are replacing more traditional broadcast services, resulting in increased demand for network resources. Multicast techniques may be used to improve bandwidth utilisation by servicing one or more clients with each network stream. In this paper, we re-evaluate the performance of the existing multicast Patching technique. In particular, we build on theoriginal server-centric performance evaluation to examine the effect of Patching on an entire network. We also propose a novel approach to the implementation ofPatching using standardised transport control (RTSP) and multicast control (MLD) protocols. Our results highlight how network distance from the origin servereffects the improvements gained by using Patching.

On-Demand Multicast Streaming Using Collaborative Prefix Caching

Increasing mass-market acceptance of on-demand streaming services motivates us to seek new innovations in the way we deliver media content over networks. An architecture is proposed in which edge-resources in a peer-to-peer network assist in the provision of fully-interactive on-demand streaming services based on multicast. This approach represents a synergy between multicast batching, proxy prefix caching and collaborative caching of media content. The approach differs from other work in its use of long-term caching of content across streaming sessions. The proposed approach has been evaluated using a highly detailed network simulation that models real-world network (IPv6) and streaming (RTSP) protocols and the Pastry overlay network. Results demonstrate that substantial reductions in server bandwidth can be achieved with low client storage and bandwidth overhead.

Application Support for Virtual GPGPUs in Grid Infrastructures

Scientific computing on grid infrastructures has historically focused on processing vast workloads of independent single-core CPU jobs. Limitations of this approach, however, have motivated a shift towards parallel computing using message passing, multi-core CPUs and computational accelerators, including GPGPUs in particular. Application support for the use of GPGPUs in existing grid infrastructures is still lacking. A model is proposed for the orchestration of GPGPU-enabled applications based on commonly used frameworks such as CUDA and OpenCL. The model makes use of recent advances in remote GPGPU virtualisation, making it possible for an application to access GPGPUs installed on remote hosts. Each physical GPGPU is isolated, creating a pool of virtual GPGPUs that can be allocated independently to jobs. A proof-of-concept Grid supporting virtual GPGPUs has been implemented and tested. It will be shown that users can be provided with a simple yet flexible and powerful mechanism for specifying GPGPU requirements. Furthermore, vGPGPU provision can be fully integrated with existing grid middleware and services. Performance results suggest that improved resource utilisation can compensate for the overhead of remote GPGPU access.

GPGPU Virtualisation with Multi-API Support Using Containers

Virtualisation of GPGPUs using PCI-Passthrough is limited to costly specialised hardware. API-Interception provides an alternative software-based approach to GPGPU virtualisation that has been shown to provide good performance and increased utilisation on High Performance and High Throughput Computing systems. Furthermore, user applications can transparently access many non-local GPGPU resources. However, current API-Interception implementations have either limited batch system support or none at all. This paper introduces a new multi-component system that supports multiple API-Interception implementations on several batch systems. The system consists of: (a) a factory component that produces lightweight Linux Containers using Docker, where each Container supports one or more API-Interception implementations and controls a single GPGPU; (b) a registry service that manages the Container resources; and, (c) a set of plugin scripts that bridge between the batch system and the registry. This paper also evaluates the performance of benchmarks on the prototype.