rmsd results of amd_is_logical

[__k]

Hello again,
can anyone explain how can a low rmsd be of less quality?
amd_is_logical's testing was said to be less accurate. why?

TIA Kobi

[Brian the Roman]

It's not that they're less accurate - they're just as good as anyone else's. However, since he(?) has sped up his clock on one of his accounts, his machines on that account will give up sooner when they run into collisions. By doing that he gives up the opportunity of finding that somewhat elusive better fold. But, he gains the advantage of moving on to new folds quicker. It comes down to the quality (takes more time) and quantity tradeoff.

The only way to see who is on the best side of the balance is to wait for other people in the group to start reaching gen 250. When that occurs, their results will be either better, roughly the same or worse. If they're better then he is hurting himself by giving up too soon. If roughly the same or worse then we're wasting too much time working on folds that don't produce better results. It's actually a VERY interesting experiment and usefull to the project also since it offers objective guidance on the quantity / quality balance.

It hasn't gone un-noticed; Howard is, I believe, looking at his results and will incorporate what he finds into a later release.

Once again this is something that could be set dynamically by the server.

ms

[Brian the Fist]

Originally posted by Brian the Roman

Once again this is something that could be set dynamically by the server.

ms

Perhaps, but only for people with a permanent 'net conenction :(

[Brian the Roman]

I was thinking more along the lines of the client keeps following the last instructions sent by the server until it uploads its next set of results at which time it also picks up new parameters from the server. People without permanent net connections would obviously do it less frequently than the rest of us, but it should be better than nothing.

ms

[AMD_is_logical]

Originally posted by Brian the Roman
It's not that they're less accurate - they're just as good as anyone else's. However, since he(?) has sped up his clock on one of his accounts, his machines on that account will give up sooner when they run into collisions. To clarify, I am using two XP1800 on each account (four nodes total). The "AMD is logical" account has the timeout set to about 4 seconds. The "AMD beta test account" has the timeout set to about 2 seconds. I started crunching to the latter a few days later, but with the shorter timeout those nodes are crunching faster. (But nowhere near twice as fast, because the minimize/trajectory overhead doesn't get sped up.)

Each node has produced 3 full sets of 250 generations (12 total), and should produce many more by the time beta4 is over.

I don't expect to get the best structure in the end. However, I am getting very good structures very quickly, and those structures could be used as the starting point for some slower, more carefull crunching with a longer timeout.