This is a follow-up to a previous post.
VENUE: Niels Bohr Institute, Blegdamsvej 17, 2100, Copenhagen
TIME: May 12, 2009, 14:00
14:10 – 14:40, “Cloud and Virtualization to Support Grid Computing”,
Ass. Professor Ruben S. Montero , DSA research group, Universidad de Complutense, Madrid
14:40 – 15:00, “Perspectives for clouds in scientific computing”, Dr.
Frederik Orellana, Niels Bohr Institute, University of Copenhagen
15:00 – 15:30, “The NDGF cloud”, Dr. Michael Grønager, Nordic Datagrid
16:00 – 16:30, “The Sun Cloud strategy”, Dr. Torben Kling-Petersen, SUN Microsystems
16:30 – 17:00, “A joint Nordic green computing facility on Iceland”,
Rene Belsø, Danish Center for Scientific Computing
The event was well attended – and in the opinion of the organizers a big success. We had interesting contributions on several aspects of cloud computing. The program includes links to slides (SUN’s slides are missing).
The discussions circled around whether or not cloud computing has anything to offer to HPC. It was agreed that cloud computing has two aspects:
1) fabric management or the automation of resource provisioning, i.e. streamlining the process of configuring and booting up machines – virtual or not
2) the interface offered to the outside world – currently, this de-facto means the EC2 REST interface
My standpoint was that since, at least in Denmark, grid interfaces have not really caught on with HPC users, it could be interesting to see if cloud interfaces can do better. They arguably put more work on the user who just want a regular batch system with some Linux boxes as worker nodes. My question is: does such a user exist? Most compiled applications depend on system libraries and therefore a prime reason why HPC users are so happy with their local systems is that on these systems their applications actually run (they were compiled there). A free choice of platform on any resource, local or remote, would eliminate this problem and this is what cloud computing can potentially offer.
Thus, it was more or less agreed that a crucial premise of the discussion is: who, if any, are the HPC users we would like to lure away from their local queuing interfaces and onto a more distributed platform? This question caused some debate