To answer this question, I at first have to argue, why scientific computing like ATLAS Computing should profit from Cloud Computing with virtual machines in contrast to e.g. the LHC Computing Grid (LCG) approach.
In computer centres of the LCG the need for software maintenance (operating systems and job systems) is very big, because unstable software and particularly user mistakes often lead to crashed systems. Because of this, the personnel effort and costs for administration and maintenance in the LCG is big.
So, what are the advantages of working with guest operating systems on virtual machines (VMs)?
- The user does not touch the host operating system.
- Different VMs on the same hardware work independently.
- The user can work as root without worry.
In other words:
- The user is not able to damage the host OS (bad intentions excluded).
- If one VM crashes, the other VMs on the same hardware proceed working.
- The user may “destroy” one VM, but he can just relaunch it (takes advantage of the machine image concept)
So, when the virtual machine monitor (the software layer providing the virtualization) runs stable, there ideally should be nothing, that could crash the host OS. From the computer centre staff’s point of view, the software (the host OS) then nearly is maintenance free. With virtual machines in a cloud, computer centres would almost only have to care about the hardware, which would reduce personnel effort and expenses!
You perhaps ask yourself, if there is a loss of performance, when hardware stressing jobs run on virtual machines. The solution is the open source paravirtualization Xen:
«Through paravirtualization, Xen can achieve high performance [...] which is notoriously uncooperative with traditional virtualization techniques.»
This requires the guest operating system to be modified in a special way, which is not a big problem, as we can see using the example of Amazon Web Services. AWS uses Xen virtualization for EC2 and because of this, there the guest operating systems need a so-called «Xenified kernel».
Hence, there are various advantages, Cloud Computing with virtual machines brings along.
The first reason is that a reliable technical infrastructure to give Cloud Computing a try is currently only commercially available. The leading company in this sector is Amazon with its Amazon Web Services. Additionally, Amazon really is interested in working together with science. They sent out representatives / consultants to the Max-Planck-Institut, to consider our needs and to work on an individual solution for special applications in science. So, for Cloud Computing research purposes, AWS is ideal.
When calculating and comparing the costs for computing with AWS and with own hardware, it is obvious, that AWS is more expensive. It is definitely excluded and not the intention to e.g. move ATLAS Computing completely to AWS.
But there is another serious advantage, Amazon Web Services for ATLAS Computing brings along: imagine, there is a deadline and you have to get e.g. simulation results until a fixed point in time. Additionally imagine, that your classical computing possibilities run out of capacity and it is foreseeable, that computing will not finish. Then it might be of very big value for you, to occupy a huge amount of instances at EC2, so that you can meet the deadline.
The described case is exactly the case AWS is designed for. It offers flexible, scalable and elastic computing infrastructure, so that a company or even science gains planning reliability. Remember, that the AWS user only pays, what he uses!
In summary, AWS offers a very good base to develop Cloud Computing for scientific applications. AWS is promising in respect of missing computing capacity. Those peaks of desired computing power can be satisfied using AWS for a short time. The resulting cost may be less important than meeting a deadline.
Let me summarize:
- Cloud Computing using virtual machines would reduce personnel effort and cost.
- Computing with AWS is more expensive than with own hardware.
- AWS offers a great possibility to satisfy peaks of desired computing power
These statements implicate, that the optimal solution would be an own computing cloud for the price of own hardware in combination with a standard Cloud Computing API that is able to control the own cloud and additionally e.g. AWS. Then the user is able to decide, in which cloud his jobs should run. Switching between clouds becomes very easy, resulting in a very convenient solution to balance out peaks of desired computing power.
This would be a very profitable mixture and research on this topic is exactly the main motivation of the AWSAC project.
