I'm not so sure I see the great usefulness of this attack.

I've taken a cursory glance at Mr. Kocher's paper on-line and what it comes down to essentially, if I undestand it correctly, is that you need to be as sure of the timing as you can be.

Now, on a distributed system, you can't measure those timings, because any latency could come from the originating computer, the links in the middle or any combination of them.

Also precise timings can be limited by fluctuating load averages amongst other things in a time-sharing computing environment. While this might work in a lab, with the current advances in computing speed, the differences between a fast and a slow calculation can easily be opaqued by network lag.

Am I missing something, or does this attack only work in a lab?

The more timing noise between the attacker and the target, the harder it is to exploit the measurements. Based on some (very rough) experiments I've set up here, I suspect the attack is easy if you're on the same computer (and measure CPU load), probably feasible if you're on the same network and the host and net are unloaded, and unlikely otherwise. The attack is especially interesting against crypto tokens that are supposed to hold a secret key secret, where you can get very close and take very good timing measurements. Keep in mind also that Kocher's results are only the first cut, based on a very simple statistical model. I suspect we'll be seeing many improvements and variations over the coming months. Bottom line is that implementing good cryptosystems is a lot harder than one might think... -matt