This paper argues the for the need of an operating system on datacenters. The paper makes great points in explaining that there are now many applications for cluster computing, and that an operating system should be created to provide the ecosystem these applications need. An operating system could allow finer-grained resource sharing and data sharing across different cluster applications. It would also provide programming abstractions to provide a uniform interface to the underlying hardware, which may lead to applications becoming more compatible with one another. In addition, OSs designed for the datacenter could implement debugging and monitoring facilities that are so hard to come by with cluster computing. Overall, the paper makes its arguments very well and details exactly why the previous features would improve cluster computing.
One comment I have to make about this is that a really great point was brought up in the resource sharing concerning the role of virtualization. The authors mention that VM migration and VMs in general might simplify the scheduling over a large datacenter. I agree with this, and I think we can take it further to possibly consider proactive load balancing in datacenters. Proactive load balancing has already proven itself to be very effective in the HPC world, so I think it would improve cluster computing quite a bit as well.
Another comment I have is that this paper doesn't really seem to address energy at all. I think one of the great advantages of having an OS for the datacenter is that now, with a broader view of scheduling, it would be possible to shut down entire racks of computers if not needed, and to preemptively to start these up if big batches of jobs are incoming. This would be another great benefit from having OSs for datacenters.
No comments:
Post a Comment