At the other extreme (as compared to the LAN-Grid) is the ideal solution a centralised computing platform to serve everyone. Think of this as the Net-Grid. It is the way we access Googe, Yahoo and eBay so far, weve used these websites for services such as email, search and auctions. Now, we need to go a step further and even have computing delivered from the centralised bank of servers. We are already seeing this emerge in the second generation of application service providers (ASPs) like Salesforce.com. So, whats different about the Net-Grid?
The Net-Grid takes the LAN-Grid to its logical conclusion if bandwidth were available, the LAN and WAN would merge into a single network, and thus the locally hosted computing platform could be moved further into the core. The edges become simple presentation and input devices, while the core houses the complexity of the computing platform.
But there are big differences. The technologies that will work on the LAN-Grid in terms of data replication will not scale to the Net-Grid. The need is for a distributed file system which can support millions of users. There are multiple options available in open-source: Andrew File System, Coda, InterMezzo and Lustre are some examples. In a sense, the challenge here is simpler than what the likes of IBM and Sun have worked on solving treating a whole host of distributed resources across a network as one through virtualisation. In the case of the Net-Grid, all the hardware can indeed be centralised into a single location. The three challenges that need to be addressed for the Net-Grid are: providing scalable computing, scalable storage, and granularised billing.
The computing grid that we are discussing is very much like the architecture available in telecom: think of the LAN-Grid as the equivalent of a PBX, while the Net-Grid is the centralised switching system. (There is also an intermediate option that of a Centrex service from the local central office think of this as an Operator-Grid, hosted by the broadband operator over a last mile which is free of opex once the capex is done.)
In terms of costs, there are two key considerations: that of the computing infrastructure and bandwidth the assumption currently made is that open-source software would be used to provide the basic set of applications for the desktop. From a user point of view, assuming we would like to stick to the aggregate Rs 700 ($15) per user per month payment for the whole ecosystem of commPuting, the allocation for the grid will be about $2. [The rest is split thus: $8 for bandwidth, $4 for the network computer rental, $1 for support and reseller margin.]
Of this $2 monthly grid payment, about half would be spent on bandwidth and data centre costs. That leave us with $1 to provide the computing and storage infrastructure. Is this doable? Let us do some back-of-the-envelope calculations. Let us start with the LAN-Grid which had a capex of $55 per user. We will get some economies of scale in pricing as we move computing to the storage. However, the real benefit that we will get is that much like a telco, we will not need to allocate that same computing power for every user. Let us assume that we design a system that assumes that a third of the users are online at the peak. This means that our capex will be a maximum of $18 per user (extrapolating from the LAN-Grid cost). Assuming that this can be amortised over a period of three years, and add financing and management costs, we will probably get to something like a total spend of about $24 over three years, or about $0.67 per month. This gives us an operating margin of $0.33 per user per month or about $4 per user per annum. So, in theory, it should be able to build, own and operate such a public computing grid.
TECH TALK CommPuting Grid+T