die erste Adresse für Distributed Computing
Unterstütze Rechenkraft.net e.V. |
Mit Suchkraft hier suchen:
graeme wrote:Our focus this year has been on the computational algorithms. My postdoc, Lijun Xu, has implemented several strategies to increase the simulation time, accuracy, and system size scaling.
graeme wrote:One improvement involves taking all the important processes found at one step, and sending them back out as initial guesses at the next step. This helps a lot if the system is large enough that the processes are confined to a local region. When a process is selected at one step, any remote processes will be unchanged and returned to the server very quickly as a new valid process at the subsequent step.
graeme wrote:A second improvement also makes use of this locality. We are testing if we can focus new searches at the current steps in the area in which a process was selected at the previous step. If this works as we expect, it will lead to linear scaling in the cost with respect to system size.
graeme wrote:One final improvement is how we deal with fast transitions between known states. This used to be the main reason that the server would not respond to clients - it was busy simulating the dynamics between states that it already knows about. Now, we apply a statistical model to choose a processes which leads to a new (unknown state). This has significantly reduced the server dead time.
grame wrote:Lijun is writing up a paper on these methods, and we'll have it out this summer.
graeme wrote:Largely based upon the results from eon, we have received a NSF grant for taking this project further.
One of our long-term objectives is to use quantum mechanics to evaluate the interactions between atoms. These calculations are much more expensive that what we are doing, so that only the fastest machines with enough memory could participate. But, as computers get faster, this will eventually be possible. When we get this working, it will mean that we can simulate more interesting systems, including chemical reactions at surfaces, and the computational design of catalysts.
Patrick1984 wrote:Bin dabei!
Aber XtremeSystems hat nen recht hohen output, da müssen mehr mit ran.
Also los gehts! Kampflos geben wir doch nicht auf.
Schnizzel wrote:Gibts da irgendsoeinen Richtwert wieviel Instanzen man laufen lassen soll? Hab auf meinem C2D immer nur 2 Instanzen laufen gehabt, dachte halt wie bei jedem anderen Projekt eben auch.
Wenn ich das mit deinem Verhalten grob umrechne, sollten 12-16 Instanzen bei mir drin sein oder?
Ist das effektiv oder wirds irgendwann zuviel?
Users browsing this forum: No registered users and 2 guests