As we all know Cloud computing has become more of a buzz word than anything else these days. Different people define it in many different ways. But we can certainly see some distinguishing features of it. I think cloud computing encapsulates important features of virtualization, utility computing, Software as a service, grid computing, etc..[1] Other than that I don't think it is a good idea to try to define it.
From the academic point-of-view, what will this give to the scientific computing and what are the challenges?
The first question is mostly about the programming model. As we know most of the scientific applications today are exploiting the functionalities of MPI. But will these applications, written using MPI, work well in cloud computing environments? The overhead of virtualization can be a bottleneck of MPI, within clouds. But certainly there are [2,3]/will be approaches to improve the latency of virtualization. Then the question will be how much effective, usable they will be within those data centers.
Professor Dennis Gannon proposes a model[4] around three key characteristics. He envisions a hybrid programming paradigm which will integrate the best of "Software-as-a-service", "OS Virtualization" and "Parallel Frameworks. As we already know Google already uses MapReduces (parallel framework) and their AppEngine(SaaS) within their environment.
A primitive question that might come up today with scientists is that "where should I run my applications? Is it on grids or in clouds?". One of the important factors in this decision is the cost associated with clouds if we use existing commercial cloud vendors. This is in addition to the virtualization overhead on MPI and data transfer/maintenance problems within clouds. Ewa Deelman[5] had done a good study on this with Amazon. One of the important outcomes in that study is the cost of data transfers. It is a good idea to know data transfer patterns because it is much cheaper to re-use data in the cloud compared to moving them everytime those are required.
But the long term question is much more different than that. Given a set of hardware how do we decided whether to put a Grid like infrastructure or a cloud like infrastructure? What will be the determining factors for me to select grids vs cloud? I don't think there is enough discussion happened on this.
(Kudos to my lab mates, especially Suresh, for some enlightening discussion on above topics during our lab coffee time)
[1] : http://www.appistry.com/cloud-info-center/index.html
[2] : Operating System Support for Virtual Machines Samuel T. King, George W. Dunlap, Peter M. Chen Proceedings of the 2003 USENIX Technical Conference - 2003
[3] : Xen and the art of virtualization P. Barham, B. Dragovic, K. Fraser, S. Hand, T. Harris, A. Ho, R. Neugebauer, I. Pratt, A. Warfield. Proceedings of the nineteenth ACM symposium on Operating systems principles 2003, Pages: 164 - 177
[4] :The Computational Data Center – A Science Cloud. Dennis Gannon.
[5] : The Cost of Doing Science on the Cloud: The Montage Example Ewa Deelman, Gurmeet Singh, Miron Livny, Bruce Berriman, John Good
Read more...
