Computational Power

Amazon

Marco Andreetto is wondering if it would make sense to use this service: [[1]]. Please post your comments here.

Pietro's calculations

When we buy a machine we spend approximately 2K$. The machine is used for approximately 2 years, so it costs us 1000$ per year. At 10c per hour of use, we could buy 10,000 hours per year of Amazon's machines, equivalent to 2500 hours on one of our machines (Amazon's machines are roughly 1/4 as fast as ours: 1 CPU instead of 2 and approximately 1/2 the CPU MHz). Questions: how many hours is one of our machines used? There are 356x24~10^4 hours per year and we probably use a machine 10% of the time -- approximately 1000 hours... this calculation would suggest buying time on Amazon instead...

Alex Comments

My 2 cents ...

I'm not sure i completely understand -- but it seems that we somehow upload images of our work to this site and run them remotely?

I think this is a bad idea, although have to say I don't understand the details of it ... Concerns: (1) Often I use specialized software, requiring specific libraries, etc. I forsee lots of time configuring and getting everything to work on their computers. (2) Often I monitor the progress of my experiments using a GUI -- will these be easy with this system? Very rarely do I just run a program and wait a day for results without checking on them -- I would say this situation occurs about 10% of the time, and would require one to upload code each time there is a tweak? Seems very roundabout/time expensive/tedious to me... would it take a long time for images to pop up?

Perhaps though using this system during crunch time might be advisable ... maybe a combination of local computers with this ...

Testing the system

The service is not available now. Only a limited number of users are allowed for the current beta version of Amazon EC2. We are currently in the queue.

Sun Alternative

There is a possible alternative from SUN [2]. The cost of the machine is 1$/(Hour-CPU). The system requires SunOS compatible applications and it seems similar to the CS grid here at Caltech, i.e. the jobs are executed in batch and each application should be self-contained.