Yesterday in one of the sessions in the Dealing With Technology Risk Waters Day (http://db.riskwaters.com/public/showPage.html?page=dwt05_newyork), an open discussion about GRID's started.
Here are a few important observations.
Scott Marcar JPChase has for the trading floor a GRID of 2500-3000 CPU's, he found that trying to use idle PC's didn’t work, he created a special back bone just for his GRID off of which he has IBM Blades and Genera boxes. He has managed to reach an internal running cost of a bit below $1 per CPU Hour, he is not charging his respective departments yet and feels he could get it to below 50 cents. Applications/Businesses utilizing his GRIDs are Market Risk, Risk, FX, and FX Options.
Marc Baumslag from Merrill first emphasized not to confuse GRID Computing with Parallel Computing. They tested and built a few GRIDs’ for their Risk and Credit Risk and found that they are simply not worth it. In general he feels GRID technology is an efficiency measure one should somehow harness as probably most PC's have at least a 50% CPU idle time and that is simply a waste. So in concept GRID should create an efficiency and thus save money. This is why he is still testing, though with no real positive finding. He did point out that a GRID would only work if you have an application or a process, which you can easily, split and distribute. For complex risk calculations etc. one must go with parallel computing, which he indeed does successfully.
Joseph Panfil CME, will not touch a GRID!. Needs reliability and speed, hates slowness and latency. His electronic trading is heavily dependant on Parallel Computing.
The more I meet and hear from people who actually put time and money and applications into making GRID work the more I feel the following;
The investment in getting a GRID to work, and to generate an efficiency and maybe even a cost save is enormous. Thus it is probably only worth it if you have an enormous amount to save.
In order for a GRID to work, don’t fool yourself, you will need to create a NEW network, probably an Infinband band, and you will probably need to rewrite some of your apps and not all your apps or businesses will be able to benefit.
PC's with idle CPU's are not necessarily a gold mine and maybe if you really need to utilize a GRID, The Sun $1 per CPU hour is not such a bad proposition after all.