HPCS 2012 Paper Presentations – Cloud and GPU



Topic Area: GPU

Alexandru Patriciu, Assistant Professor, ECE Department, McMaster University

CUDA Accelerated Needle Insertions in Deformable Tissue

We present a stiff needle – deformable tissue interaction model. The needle – tissue interaction is modeled using fundamental boundaries. The implementation is accelerated using NVidia CUDA. The proposed model was integrated in a custom software that loads DICOM images, generate the deformable model, and simulates different insertion strategies.



Topic Area: Clouds

Mahfuzur Rahman, University of Manitoba

Hybrid Resource Provisioning for Clouds
Mahfuzur Rahman will speak about his work on hybrid provisioning in Compute Clouds. Provisioning is the process of deciding which virtual machines will be run on which physical machines in the cloud facility. The most common approach to provisioning is to do this statically based on pre-agreed upon expectations of need described as service level agreements (SLAs). While this has not run-time overhead it does not permit adaptation to changing loads. A more recent approach uses live VM migration to do provisioning on the fly but with potentially significant overhead. Hybrid provisioning seeks to “tune” a good initial static placement using dynamic techniques. Mahfuzur will discuss the status of his prototype implementation based on his extension of the Eucalyptus open source cloud system, his plans for evaluation of specific hybrid techniques using CloudSim and the possibility of using the effectiveness of the dynamic tuning to trigger a static re-provisioning to ensure future performance efficiency.

Samer Al-Kisawny, PhD Candidate, Department of Electrical and Computer Engineering, University of British Columbia

Avoiding the Disk Bottleneck in Deducplicated VM Transfer
Data deduplication is a commonly adopted optimization in the cloud infrastructure to reduce the cost of storing and transferring virtual machines (VM). Deduplication efficiently identifies and eliminates similarity across VM image files leading to a reduction in the amount of data that need to be stored and/or transferred. This reduction in data size is significant: recent studies show that similarity across virtual machines can be as high as 96%. Moving a group of VM images within, or across, data centers is a frequent operation: it enables application migration, new application deployment, as well as backup and maintenance operations. While deduplication reduces the overall size of a group of VM images their efficient transfer and, particularly, their re-incarnation at the destination site are still costly operations. This is mainly due to three main reasons: high deduplication overhead, transfer metadata high memory footprint, and the non-sequential disk access.This paper presents three optimization techniques to accelerate the migration of a group of VM images, namely: distributed deduplication, distributed metadata management, and optimized IO operations. The paper presents the system design; model based evaluation as well as a prototype evaluation with real workloads. Our evaluation shows that the proposed techniques can bring significant performance gains, up to 4 times faster image transfer time compared to the current alternative approaches.


Individual Presentations

Intended Audience 

General Audience: Faculty, Staff, Students

Session Duration

60 minutes

Session Schedule

Thursday, May 3rd
12:30 – 1:30 p.m.
Canfor Room (1600)


Rob Simmonds, WestGrid Chief Technology Officer