Friday, November 30, 2007

The 4th fastest supercomputer, continued

In continuation to the email from Viral Shah on my blog entry titled Interpreting the 4th fastest supercomputer in the world, Subhomoy Bhattacharjee got Viral to write an article in Financial Express on the subject. In addition, Business World has a story describing the `Eka' project.

2 comments:

  1. Some comments on the Viral Shah article:

    (1) He in some sense disparages the use of commodity hardware assembled clusters as opposed to custom architectures like the IBM, Cray, Altix etc. I respectfully and vehemently disagree. In my experience with clusters I've thought that the proliferation of custom hardware only hinders the development of generic good quality code. Each code seems to be hardcoded to the quirks of each hardware implementation (for the initiated FFT libraries anyone?)

    (2) The cost / MFLOP of the custom made systems tends to be higher too. Especially relevant for systems in a nation such as India. I concede that the specialized hardware might have some advantages in terms of speed, elegance and ease of maintenance. But are they sufficient to justify the added monetary outlay? More so when we factor in the relatively cheap and abundant “manpower” in India. (You’d be surprised at how much helpdesk and troubleshooting goes into one of these systems!)

    (3) Mass production systems from cheap components always get the economy of scale upper hand. To wit the near demise of the Cray over the decades although admittedly it did make a comeback over the last few years.

    (4) Ultimately, in my opinion, if the sole objective is “national-pride” or making it to some top-nnn list then custom hardware makes sense; but if we want machines to solve some economically useful problem I’d put my money on the commodity hardware!

    ReplyDelete
  2. Rahul,

    Commodity hardware makes sense if you want to make it to the top500 list, because thats when you want to minimize cost/MFLOP.

    If you want to solve real problems, then its not just the cost of the computer, but its the cost of developing the software, and the metric you want to then use is cost/science done.

    ReplyDelete

Please note: Comments are moderated; I will delete comments that misbehave. The rules are as follows. Only civilised conversation is permitted on this blog. Criticising me is perfectly okay; uncivilised language is not. I delete any comment which is spam, has personal attacks against anyone, or uses foul language.