EVENTO
An Approach for Definition and Placement of Virtual Machines for High Performance Applications on Clouds
Tipo de evento: Seminário de Avaliação - Série A
The efficient deployment of HPC (High Performance Computing) applications on Clouds offers many challenges, in particular, for communication-intensive applications. The consensus of the relevant literature is that machine virtualization introduces performance overheads due to increased message latency and variability, so that tightly coupled applications execute with much lower performance as they are sensitive to these performance metrics. There are several lines of research to mitigate this problem, such as: a) virtualizing better interconnects like Infiniband, b) enhancing and selecting virtualization technologies based on resource requirements, c) development of VM placement strategies, d) modeling of resource utilization and application classification, and this work is supported on the last two approaches. Current works on VM placement that seek to optimize performance by using application classification have one or more of the following limitations: a) assume that the application context is constrained to a single VM, disregarding virtual clustering; b) validate heuristics with workload simulations, overlooking resource contention; c) support the provisioning process found in IaaS (Infrastructure as a Service), where the user selects the VM profile (VM provisioning), after which the provider maps the requested resources to physical machines (resource provisioning). The last limitation disallows the possibility of choosing the virtual cluster topology, which can produce substantial impact on performance.My work aims to serve PaaS (Platform as a Service) e-science Clouds, and assumes that the provider is capable of VM provisioning decisions, i.e. the user need only request a number of processor cores for a known application, after which the provider chooses the VM profiles and schedules the VMs on the infrastructure. My main contribution is a methodology with the following features: a) the representation of VM placement and virtual clusters through the placement of virtual cores; b) a software for systematic execution and obtainment of metrics under different configurations; c) performance analysis aware of core placement, resource contention and resource usage patterns; d) modeling and prediction of performance by feature extraction using Canonical Correlation Analysis on kernel metrics, according to the Dwarf classification. The objectives of this work are to detect and model sources of resource contention that arise from virtual clustering, and to produce deployment guidelines according to the required application and number of processes.
Data Início: 24/01/2014 Hora: 10:00 Data Fim: 24/01/2014 Hora: 13:00
Local: LNCC - Laboratório Nacional de Computação Ciêntifica - Auditorio A
Aluno: Giacomo Victor Mc Evoy Valenzano - Laboratório Nacional de Computação Científica - Bull Atos
Orientador: Bruno Richard Schulze - Laboratório Nacional de Computação Científica - LNCC
Participante Banca Examinadora: Antônio Roberto Mury - Laboratório Nacional de Computação Científica - LNCC Bruno Richard Schulze - Laboratório Nacional de Computação Científica - LNCC José Neuman de Souza - -
