I read thru the technical report from UC Berkeley, Above the Clouds: A Berkeley View of Cloud Computing with interest. My analysis:
- As an undergrad work on cloud computing, the paper gets an A+. But as a position paper from eminent academics, I can only give a C-. Granted it correctly identifies many of the trends and obstacles. But that material is widely available !
- With a title “A Berkeley view of cloud computing” the report misses the point. “A Berkeley observation…” is more like it – view requires original thinking and interpolation, which the report lacks.
- The authors got some of the essentials of Cloud Computing right viz: infinite capacity, no up-front commitment and pay as you go.
- The three classes viz: amazon , Microsoft and the Google model is interesting. But there are more in-between.
- They have some good points on the cost advantage of power et al and leveraging that aspect by building datacenters at the appropriate locations.
- The new application models viz. analytics, parallel batch processing, compute-intensive desktop applications and so forth are excellent observations.
- They have done some good work in characterizing elasticity. Pages 10 and 11 are good read – the models are very simplistic, though.
- They also have done a good job in showing the economies of scale that can be achieved by a cloud computing infrastructure.
- I like their assertion that “there are no fundamental obstacle to make cloud-computing environments secure as well as compliant to data and processing rules. Declarative policy and enforcement thereof is my answer.
- They have correctly identified scalable storage as one of the bottlenecks. The BigTable(Google), Dymo(AMZ) and Cassandra(facebook) all are solutions for the challenge.
- But, they got the model wrong ! The essentials of Utility Computing is the consumption model not the payment model. No doubt the pay-as-you-go model is attractive to startups, but the payment model is the second order effect. For enterprises and other organizations, the value proposition is the elasticity and the just-in-time availability of resources. Even for startups the pay as you go is attractive but elasticity is much more important.
- Argument about increase in performance and resultant cost reduction. This just Moore’s law and it is achievable within IT environments as well as a cloud computing space. I think computers are on a 5 year amortization schedule and depreciation. And a refresh can be done – with associated efficiency whether they are a cloud provider or an IT shop.
- I think the major disconnect in the paper is the basic definition of a cloud as public. The artificial separation of public/private clouds and the focus on payment were the two areas where their definition has gone awry. Cloud is an architectural artifact and a business model of computing. But clouds are clouds – internal or external, public or private. The internal vs. external is only a spatial artifact – which side of the firewall. Not worth a demarcation when we talk about the domain of cloud computing. Which side of the internet (firewall) does the cloud infrastructure lie, should not be the criteria. By their definition, they have disenfranchised the whole set of clouds inside organizations. The internal-external cloud integration across data, management, policy and compute planes is an important topic which this model conveniently skips. Also as I mentioned earlier, utility is the consumption not a payment model. A big organization can have a cloud computing infrastructure and it’s business units can leverage the elasticity – no need for a credit card, a charge back model will do.
- I really didn’t get the “statistical multiplexing” they mention a few times. What exactly is this and what is the relevance ? Just a buzz word to jazz up the paper ?
- I literally got lost in their characterization of DDoS attack and the cost models there of on P.15. Really convoluted and it does not change for traditional vs. cloud. They found a break-even point for DDoS attack based on very slim assumptions.
- I do not think the data transfer bottleneck, as described in the paper (P.16), is an issue. Who is going to transfer 10TB of data routinely for cloud processing ? Looks like a force fit for some calculations done by someone.
- The report has no useful tables or equations. Equations 1 and 2 (which are the same, btw) are not right – in thesense that the datacenter cost includes the utilization and I do not think we need to accommodate for it additionally.
- I am sorry to say all the cost models and the equations look forced and very unnatural. Even the assertion of 1/5 to 1/7 cost advantage of a datacenter is at best questionable.No value what so ever – sorry folks