The reality of Grid computing

The reality of Grid computing

Summary: David Berlind did this excellent interview with Wolfgang Gentzsch, who is one of the pioneers of grid technology,anddispels some of the myths of grid computing. In the past, I've tried to dispel some of the grid mystique myself but it's been a challenge to get the point across.

TOPICS: Hardware

David Berlind did this excellent interview with Wolfgang Gentzsch, who is one of the pioneers of grid technology,anddispels some of the myths of grid computing. In the past, I've tried to dispel some of the grid mystique myself but it's been a challenge to get the point across. Now, even Mr. Gentzsch makes it clear that the grand vision of a grid over the Internet is just a bunch of hype. He goes on to say that grid computing may have a role between organizationssuch as a hospital that leases high speed private lines to link with other hospitalsthatmay have idle CPU clock cycles to spare.Here's my question: The other hospital may have idle bed sheets, idle microwave ovens, idle beds, idle x-ray machines, and idle operating tables. Does that mean it makes sense to fly the patients and doctors around via private helicopters between the hospitals?

The first thing we should be clear about is that hospitals are not in need of greater processing power. Hospitals are mostly in need of better infrastructure such as wireless LANs, better software, better systems integration and implementation, and better security. Even if a hospital (or any organization) needed more computing power, it would be much cheaper to build a cheap super computer using commodity AMD- or Intel-based hardware than to pay through the nose for monthly OC3 service -- justas it isn't feasible to lease your own helicopter fleetin order to share some physical resources. Even if budget was no object, an OC3 circuit still has less than 1 percent of the throughput of the type of Cisco WDM optical gear mentioned in David's article, so it will be very limited to the types of grid applications that can be run. One of the biggest fallacies of the logic behind the "need" for grid computing is that processing power is somehow scarce and/or expensive. That couldn't be further from the truth because processors are dirt cheap. Building a bank of AMD- or Intel-based PCs into a super computer is, relatively, very cheap and very practical, not to mention that you have the luxury of cheap gigabit interconnects in the form of CAT5 or Multimode Fiber cables within the cluster. The bottom line is, the speed and cost of a localized cluster will always be an order of magnitude superior to a wide area cluster.

Finally, the biggest obstacleto grid computing -- which no one hasanswered convincingly -- is security. It's one thing to store or transmit encrypted content on an untrusted medium, and the solution is relatively simple and trivial. It is something else to process transactions on an untrusted processor; at some point, the data must be unencrypted before it can be processed. I have yet to see a grid that can take in encrypted data that it has no way of decrypting, process it, and then output encrypted data that only the source of the original job can decrypt.

Topic: Hardware

Kick off your day with ZDNet's daily email newsletter. It's the freshest tech news and opinion, served hot. Get it.


Log in or register to join the discussion
  • Vertical or Horizontal?

    Grid technologies allow for horizontal scaling of computer resources. This is NOT so good if your application cannot be parallelized. IBM and its Power 5 chip will also lessen the need for Grids, since you can activate CPUs as you need them. Grid's can only make sense for the "client" infrastucture (not server), and there are many pitfalls - like people shutting off or rebooting their client machines. Also, lack of a common "namespace" i.e. common entreprise-wide directory structure, hobbles the Grid by making it transfer data back and forth, and routing it to the "right" nodes.
    Roger Ramjet
    • Even worse than that

      Roger, you make some excellent points. Grid is by far the slowest form of horizontal scaling because of the bandwidth constraints. With a localized cluster, you have the luxury of relatively cheap 10 gigabit interconnects where as it would be out of the question to lease a one gigabit metropolitan area network.

      Even with multi-gigabit interconnects; horizontal scaling still has a serious problem scaling on many applications.
  • grid computing

    while i tend to agree that the idea of somehow obtaining cycle on the fly for sensitive apps, for a business application might not be feasible, i am reasonable convinced that you can create a system where you are able to run specific classes of applications on a wide network of machines...

    i envision a system where you might have 20K-30K servers across the internet, and you ahve the ability to run certain types of apps on these systems. the network/systems are able to be configured on the fly to allow you the user to run the given app on the network...

    with this approach you can allow a small design team, to be able to run large/complex eda apps for their design, using a system that they couldn't normally be able to afford to access...


    • It's called SETI

      I think that even one of the "free" ISP providers offered a service where their service "free loaders" had to give up CPU cycles that the ISP could sell. The problem is, anytime you ask someone to pay dollars per instructions when you could buy the CPUs for much less, it usually doesn't fly.

      The bottom line is, the types of applications that you could run on this type of low bandwidth Grid is extremely limited. Most Grids are local Grids for a good reason, and even their function is very limited.