Skip to main content
Department of Information Technology

Performance bottlenecks when running in parallel on Cloud Virtual Machines

For certain classes of scientific and technical computing the cloud may offer easily accessible, scalable, and affordable gigantic compute power. A power that for these classes may lead to a step change in model and analysis complexity compared to what is feasible with dedicated clusters and similar networked solutions.

However, it is also known that seemingly embarrassingly parallel tasks may scale poorly on cloud virtual machines and this project work aims at explaining
such performance and at detailed analysis of the interplay between problem size, computing effort, communication, the ratio pCPU/vCPU, and other parameters.

The goal is to find the main parameters describing the difference in performance between a virtual machine with a certain number of so called instances, each with a certain number of virtual CPUs, and the corresponding server equipped with the same number and type of physical units.

Any theoretical findings will be supported by HPC benchmark examples run on a commercial cloud, and if time allows, one or two commercial SWs applied to scalable industry applications will also be analyzed. It appears standard benchmarks are based on the Conjugate Gradient method for communication performance, FFTs for both communication and computation performance, whereas any embarrassingly parallel task with a minimum of sequential steps can be used to assess computation performance.

The task can be suitably split between two students and the main elements of the work are proposed to be

1) Description of the main general features of a virtual machine (VM)
2) The features of Amazon Web Services and Microsoft Azure VMs, respectively, if different
3) A literature survey on factors leading to deterioration of speed-up and other performance metrics on VMs, as compared with purely physical server configurations
4) Benchmark tests on cloud
5) Analysis of commercial applications on cloud, if time allows

Supervisors at ABB

Anders Daneryd, anders.daneryd@se.abb.com, +46 705 32 30 36
Kateryna Mishchenko, kateryna.mishchenko@se.abb.com, +46 21 34 50 55

ABB Corporate Research, 721 78, Västerås
Automation Technologies/Advanced Modelling and Control

Updated  2014-10-10 18:22:38 by Maya Neytcheva.