BLOG

While reading Nicolas Carr’s The Big Switch (pp. 37), the following paragraph struck me.

“By 1905, a writer for Engineering Magazine felt comfortable declaring that “no one would now think of planning a new plant with other than electric driving.” In short order, electric power had gone from exotic to commonplace. But one thing didn’t change. Factories continued to build their own power-supply systems on their own premises.  Few manufacturers considered buying electricity from the small central stations,  like Edison’s Pearl Street plant,  that were popping up across the country. Designed to supply lighting to local homes and shops, the central stations had neither the size nor the skill to serve the needs of big factories.  And the factory owners, having always supplied their own power, were loath to entrust such a critical function to an outsider. They knew that a glitch in power supply would bring their operations to a halt – and that a lot of glitches might well mean bankruptcy.”

With the above paragraph, Carr does a good job of paralleling the adoption of electric power in 1905 to the current state of cloud computing adoption by enterprises. Elsewhere in his book (pp. 22), Carr makes the case that “..the fostering of invention and the embrace of new technologies that result…are the consequence of economic forces that lie largely beyond our control.”

Carr is clearly making the case that similar “economic imperatives” will apply to the computing world.  In that context, I thought that it would be useful to look back at the evolution of the computing industry and see how we are progressing and understand the forces driving us towards those economic imperatives.

Economies of Scale are often touted in the context of clouds, but we should really look at progress in the context of Total Cost of Ownership (TCO).  TCO can be used to model total cost relative to benefits for a given technology.  It typically includes cost of acquiring, upgrading, deploying, operating, utilizing, maintaining, servicing and subsequently disposing a technology. The graph above roughly charts TCO for data centers over time.

My sense is that over the course of the past 40 years of computing, we have seen only marginal improvements in TCO overall, with any significant gains offset by equally significant losses due to the increased complexity and the new management burden that accompanies every new data center technology. Let me highlight some of these vicissitudes in data center TCO over the years using the chart shown above.

Way to the left, we have the early days when all of the compute elements i.e. CPU, Memory and Storage were all packaged into a box i.e. the server. This was expedient as it allowed the fastest interconnects between the various components. In this phase, initial gains in TCO came from faster hardware and servers. The increasing number, physically dispersed servers and networking all added to the complexity but for the most part we did ok with newer management tools.

Next, thanks to the phenomenon of bandwidth inversion (external FC connections with more throughput than the internal bus in a server), it became possible to have external storage connected to the server that could deliver better performance than internal storage. This facilitated shared storage infrastructure i.e. SANs, which helped boost TCO by enabling better resource utilization, improved performance and significantly better Business Continuity. At the same time, SANs also increased management complexity and caused new problems. Now we had a whole new silo in the datacenter – other than the server, to be managed and optimized. This required new skills, tools and added new costs. Additionally, all businesses or applications that shared storage could potentially be impacted if shared storage is not ideally configured. (Just as a point of reference, a basic “health check” service for SANs – a prelude to more expensive tuning services, costs approximately $400 a switch port and typically run in the neighborhood of 25,000 – $50,000 overall.)  So in the overall measurement of progress of TCO, we gained some but also lost a fair bit of ground with SANs.

This trend continues today with virtualization. Like the SAN that preceded it and revolutionized storage, Virtualization is enabling unprecedented sharing of server resources and delivering significant benefits in terms of resource utilization, consolidation and HA/DR. However, once again, we have introduced new layers of complexity and management. For example, now we have to contend with Virtual Machine Sprawl, I/O contention between applications, new security concerns etc.

We seem to be on a path that is not sustainable. Today’s data centers have become complex silos of resources (servers, storage and networks) with poor end-to-end provisioning, management and security.  What is worse is that each silo often duplicates management functionality that is also present in the other silos resulting in counterproductive results. The customer unknowingly pays for duplicated and often counterproductive management tools and services pedaled by vendors of products in each silo. This is why Cloud Computing – in my mind, represents an opportunity for a fundamental re-architecture of the data center as we know it today. Cloud Computing will require us to be application / service centric in terms of data center infrastructure and its management (as opposed to the resource-centric paradigm we are used today)….and you simply cannot deliver application-centricity (SLAs, QoS etc.) without end-to-end, visibility and control. This is where it  might be helpful to borrow on some lessons from the telecom world which also went through a similar evolution but was able to arrive at a scalable infrastructure architecture and management paradigm that delivers consistently a mission critical service i.e. voice, reliably and securely. In particular, Telecom’s FCAPS based signaling, switching and service mediation concepts could be worth looking at.

The status-quo i.e. current IT silos and layer upon layer of  management infrastructure, is inefficient and not sustainable. Virtualization and network computing have only paved the first steps, but Cloud Computing imo represents that economic imperative as Carr mentions that will necessitate a rethink of current data center architecture and management practices. With computing, we are at that threshold – just like the one Carr paints from the world of electric power in 1905, beyond which the cloud will become the only way to produce, deliver and consume ubiquitous, cheap, efficient, reliable and secure computing power.

Earlier this month, I had the opportunity to participate and present at the First International IEEE Workshop on Collaboration & Cloud Computing that was held in Groningen, Netherlands this year. Like most of you, I probably would not have heard of this conference if not for my colleague Dr. Rao Mikkilineni who chaired the workshop this year. After attending it I am convinced that this conference has some real potential and is just the kind of forum that the cloud computing industry needs in order to enhance its collective understanding, identify future directions, explore potential standards and to generally foster better collaboration between industry and academics. “Why academics?” – you say. “Isn’t the corporate world doing just fine?” Maybe so. However, it is my sense that in the rush to grab corporate market share and establish early competitive advantage, the industry as a whole might miss out on an opportunity to do things right. Let me explain.

Cloud computing represents an opportunity to usher in new levels of efficiencies by breaking down existing computing silos and enabling transparency in terms of visibility and control to users within and across datacenters.  It also provides an opportunity to design massively scalable and globally interoperable computing clouds without vendor lock-in. That said the danger I foresee, is that an industry increasingly driven by short-term thinking – due to the need to generate ROI for their investors in the near term, could lead to restrictive, non-scalable architectures that are incremental improvements over what we have today.  What is needed is some good old fashioned no-holds-barred architectural brainstorming to generate new breakthrough computing models.  This is where I believe credible organizations like the IEEE can facilitate collaboration between industry R&D Labs and academia and uniquely provide a great forum where they can work together without short term ROI pressures, to propose, deliberate and lay out a solid foundation for the industry’s future.

Personally, this year’s workshop was memorable and refreshing for at least a couple of reasons.  For one thing, I had the opportunity to visit the Netherlands – a place I hadn’t visited before. Besides that, the conference offered me the chance to re-visit the world of academic research; a world that – having spent my entire career after graduate school in the corporate world, has now almost become alien to me.  What struck me was the intimate academic environment of the workshop,  entirely devoid of the corporate agendas and marketing hype that drive almost all conferences and events in the corporate world. I found it refreshing to be able to objectively discuss current trends in Cloud Computing in this setting with other participants from various university and industry organizations, who had assembled and really had no other axe to grind than to identify long-term research themes and to collaborate on future research in the field.  The topics and themes discussed there were very interesting. Some of the key topics included:

1. Is the cloud just an XaaS stack or is there more to it?
2. Is Unified Computing simply throwing servers, network and storage in blades?
3. Are lessons from the past i.e. from telecommunications and from the Internet, relevant to enabling massively scalable and globally interoperable clouds?
4. How can infrastructure hardware vendors accelerate cloud deployment by including service enabling features
5. Is the current trend in throwing multiple OSes inside the server and including another networking abstraction layer that bridges these OSes the right model or should we look at a new network centric operating system that allows dynamic composition of distributed physical computing resources based on latency tolerance of services consuming the logical resources?
6. Can we accelerate the creation of Computing Clouds through fresh ideas such as the concept of a virtual infrastructure fabric, a management services fabric and a business services fabric?

The last topic, the concept of a “fabric” for virtual infrastructure, management and business services – essentially a SOA architecture for datacenter infrastructure,  intrigues me very much. Take a look at this paper by Pankaj Goyal if you are intrigued as well. Actually, all the papers from the conference along with an overview of the workshop are here.

Overall, some really neat stuff that had me wondering why IEEE didn’t do more with the conference in the US? Turns out the conference used to be held mostly in the US (see past conferences)…till 9/11 changed things. Apparently US visa restrictions made it hard for many to travel to the country and since then the conference has mostly lived in Europe. A shame really as I think that this IEEE workshop can be a great forum and contribute immensely to the larger discussion on cloud computing.  In addition, as much as I enjoyed the academic setting, I think that more exposure and participation from the corporate world and from the US would only serve to enrich the discussions here. I have no doubt that IEEE’s rigorous vetting and review process will ensure that the discussions are substantive and not just marketing fluff driven by corporate agendas that often pervade other industry conferences.

If you are interested, the 2010 workshop is scheduled to be held in Larissa Greece.  Who knows, maybe the next one after that will be in the US.