## Sunday, September 24, 2006

Everyone has heard of "Moore's Law", which asserts that CPUs double in roughly every 18 months. Similar phenomena are at work with memory, flash memory, disks, etc. I recently bought a pair of 400G USB disks for Rs.9,000 each - that's Rs.22.5 per G or $0.5 per G. I'm sure that some years from now, this blog entry will look incredibly dated, when hard disk capacity goes at Rs.2.25 per G! I vividly remember how roughly 12 years ago, when I was at CMIE, early design work for terabyte storage was being done and it looked most daunting. Now a terabyte is just 3 disks of Rs.9,000 each. The growth of performance of a single-CPU has not kept up with Moore's law. But two things have taken place in spectacular measure. First, holding an old level of performance fixed, the cost of a CPU which achieves that level of performance has been crashing nicely. It's easy to put a decent 32-bit CPU into just about any low-price device today. The second phenomenon has been parallel computation. The name of the game now seems to be about harnessing a large number of CPUs to get your work done. For a large class of problems, that appears to be feasible. Commercial software prices - like Oracle or Microsoft Windows - have held up remarkably well in real terms. So if one tries to configure a system today, the stable software costs loom large when compared with crashing hardware costs. But free software is creeping up into a larger and larger class of problems. It's possible to build a very powerful data centre with Linux, Apache, MySQL, and a free J2EE implementation. Such a data centre would be able to deliver transactions at mere hardware cost. ### Crashing prices of IT systems make interesting things possible Backing up your audio CDs. The `flac' file format permits lossless compression of audio CDs. So when you buy a CD, you can rip the .wav files and convert them to .flac files. As a thumb-rule, one audio CD reduces to 250 Meg or 0.25 G of .flac files. When storage costs Rs.22.5 per G, this means that for a price of Rs.8 of hard disk space, you get to keep a backup of your audio CD. When your audio CD gets scratched or broken, you would uncompress the .flac files to get back to the .wav files which can be burned into an audio CD. Imagine that. The biggest data centre in the world, through ad revenues. A few years ago, if someone had asked me whether it was possible run a million-CPU data centre supported only by advertising, I would have said that it's not possible to run a million-CPU data centre. But today, we know that Google is extremely profitable running the most complex data centre in the world, and practically all their revenues are from advertising. Other small data centres based on ad revenues? Turning to the territory more familiar to the readers of this blog, think of a brokerage firm. A brokerage firm is basically a data center interfacing between customer and exchange. If done right, using free software, it ought to be possible to do this data centre at hardware cost. If so, the cost per transaction of the brokerage firm ought to drop dramatically (see my earlier post on flat-price transactions in the brokerage industry). At some point, it'd become possible to have a brokerage firm which offers free transactions, supported only by advertising. It isn't obvious that it works, but it's worth trying. The first attempt at doing this has begun: link, link, link. Free wifi in a city. When prices of IT hardware are sufficiently low, it may be possible to have free wifi in a city. I am sure that it is possible to have free wifi within a city for low bandwidth applications like email or reading the web. I don't know how the cost of infrastructure will go up if everyone is swapping videos using bit torrent. What might happen is that when city-wide networks are built, the video-related traffic will become so great that the network becomes effectively unusable, that some kind of congestion-pricing is central to making it work correctly. This might be the case; I don't know whether this is just an unhappy scenario or a likely outcome. Easing nation-building in poor countries. I have previously written about the role of IT systems in delivering national public goods in India [link to paper]. The uniquely Indian challenge is : high volume + low transaction size. Being able to pull this off critically relies on building IT systems which deliver very low cost per transaction. A key litmus test shaping up of these questions is on the implementation of the New Pension System [link to paper]. ### Indian broadband prices are out of line In this happy environment of dropping prices of hardware and software, prices of broadband in India stick out. By now, mobile telephony in India has acceptable prices by world standards. But broadband is still way out of line. A recent blog entry says that the price of a megabit connection per month in the US has dropped to$6 (Rs.300). He quotes "Harvey's theorem": Divide your monthly bill by the speed being touted by your broadband provider" as the path to these calculations. In an Indian setting, that would be between Rs.1,000 and Rs.2,000 a month for a 512k connection from MTNL? That's between 2000-4000 INR (or $25 to$50) per megabit per month. This is between 5x to 10x costlier than in the US. I don't understand why this has to be.

China has 123 million Internet users, of which 77 million are broadband subscribers. A corresponding Indian number seems to be 25 million; broadband usage in India is as yet pitiful. I'm sure both numbers are artificially inflated - perhaps both numbers are wrong by an equal magnitude? I wonder how much of this is caused by lower Indian per capita GDP, and how much of this is caused by high prices of broadband in India.

Please note: LaTeX mathematics works. This means that if you want to say $10 you have to say \$10.