Saturday, November 29, 2008

Information Technology for the Chief Executive

At last, the english edition of my second book is available. It is somehow simplified (no technical appendix and a shorter, leaner introduction) but the body is the same as "Performance du Système d'Information".

At the same time, I just received the "Best IT book award" from the AFISI. The AFISI is the French association for engineering and information systems. I am quite honored to receive this prize when I look at the list of the previous winners.

Since I have used a self-publishing company and performed the translation myself (with the help of a professional interpreter), trying to promote this book is going to be an adventure.

Sunday, November 2, 2008

Very Cloudy Weather: will it rain on the « Cloud Computing » Parade?

This blog has been very quiet for the last few months … and will remain as such until I am finished preparing my lecture about Information Systems for Polytechnique (2009). I should have, at that point, accumulated quite some material for future posts, since I am digging quite hard into the theoretical foundations of "Information Systems".

However, I have been reading so many newspaper articles about "cloud computing" during the last few weeks that I feel like putting a few thoughts on paper. The turning point was the article from "The Economist" last week: as is rightfully noticed, too much hype will likely cause "disillusion". I am actually totally convinced that "cloud computing" is relevant to many businesses and that it represents a revolution that is coming. I have actually explained some of this in my other blog. If you are new to this, a really simple description of the benefits could the following:

  • Fault-tolerance through the implicit redundancy. This is, however, an architectural issue. It is not enough to rent your computing infrastructure from Amazon, Google or Microsoft to get this benefit. You also need to implement your information system with an architectural paradigm which takes advantage from the availability of multiple servers.
  • Super-computing performance through parallelization. The same remark applies: it is true that new techniques for data mining or real-time event processing (two examples) may be tried successfully on the cloud through a MapReduce approach; it also requires a significant amount of work if you start from your legacy application.
  • Reduced costs of operation (TCO) through the use of standardized and mass-produced units. Look at the price of the TPMC according to the type of hardware and you will get the idea. I won't dwell on this, this is explained everywhere in the newspaper articles I was mentioning.

There is an implicit warning here: although "cloud computing" is "the way to go" for many cases/enterprises, there is a learning curve and a price to be paid. Especially, it will take time and energy to move from an existing architecture to one that can be "migrated on the cloud". However, "Cloud Computing" is not the "ultimate solution" for everything and I do not believe for a second that "software is dead". Let me give three simple reasons for which one may decide to stay on the "firm ground" as opposed to ""move to the cloud":

  1. The risk of loosing privacy and control (cf. R. Stallman's reaction). There are many aspects, including legal and societal, to this issue. This is the part that is reasonably well covered in the papers (such as The Economist). It is clearly a valid point, but technology and service segmentation might alleviate this issue in the future. There may exist private clouds, secure clouds, encrypted clouds (where the encryption is managed by a third-party), etc.
  2. Latency : accessing the cloud is not instantaneous. Even if the protocols were truly optimal (and they are not, web service invocation carries a significant overload), it takes some time to access to a distant data center (the light takes 10 ms to travel 3000km, and 10ms is significant for many high-performance computations or transactions). This is why MMORPGs rely on a RCA (rich client architecture – a significant part of the work is done locally).
  3. Computational overhead: making each service invocation a web service invocation is not practical for high performance computing. There is a proper level of granularity for encapsulating a piece of computation/transaction into a cloud service. This is actually something for which there exists a significant amount of history: when companies try to develop a SOA architecture, they have to get the service granularity right. Otherwise, they "discover" that application servers cannot carry an infinite load and that they exhibit a performance overhead when compared with more traditional approaches (a RPC – remote procedure call- is more expensive than a procedure call, and a WS call is not the cheapest RPC).

Hence we get two negative answers to the following questions:

  • "Can I move all my IT onto the cloud without changing my apps and their architecture ?" (the cloud as the universal IT outsourcer). The answer is negative when performance (throughput or latency) constraints exist.
  • "Can I migrate all my apps towards a SOA architecture that will live in the cloud, made from SaaS (Software as a Service) ?". The answer is negative because some services are better produced locally. The economy of scale of the SaaS requires some form of mutualization (running a service remotely for a unique client is not a clear winner). If performance constraints prevent from breaking an application into "tiny components" (because of the cost of recomposition), many apps are too specific to be run as SaaS.

To summarize, I would propose two "axioms/theorems/conjecture" about "cloud computing":

  1. "Cloud Computing" will succeed because of the combination of two facts:
    1. The grid architecture is the best architecture to provide true scalability and superior availability (fault-tolerance)
    2. Grid architecture exhibits "economy of scale", that is, can be run more efficiently when operated on a very large scale. The more identical servers one operates, the lower the TCO per server is.
  2. "Service Oriented Architecture" (SOA) is not scale-free: there is a balance to be found between "recomposability" and performance. Hence the cloud may not host all the necessary services that any enterprise may need.

The first "theorem" could be labeled the "Google Theorem" since Google operations are an existence-proof of these two affirmations. They could be justified independently … and make perfect sense to anyone who has been in charge of running the IT department of a large-scale company. However, the same experience of being the CIO of a large company also suggests that performance and data synchronization/distribution issues will prevent quite a piece of the IS application portfolio to be moved to a cloud architecture, at least for the next few years.

 
Technorati Profile