Ho raccolto un po' di news vecchie e nuove su questo progetto:
--> Sviluppata una nuova applicazione (eOn Client 1.0) che va a sostituire la vecchia.
--> La barra con la percentuale di completamento della WU ora funziona meglio (si aggiorna ogni 2 punti percentuali).
--> E' stata attivata la compressione gzip delle WU: il download è stato diminuito di 10 volte (ora bisogna scaricate solo 14kb per WU contro i 170kb di prima), mentre l'upload è stato diminuito di circa 3 volte (ora è pari a circa 170kb).
--> Spiegazione del perché le WU non vengono replicate:
We don't have redundant computing set up because we are able to verify the correctness of work units with far less computation than re-running them.
--> Altre piattaforme (CUDA e 64bit):
The client code is not portable to CUDA. While we would like to take advantage of graphics cards, their architecture constrains what kind of calculations we can run efficiently. In order to use graphics cards effectively, we would need to be doing calculations on much larger systems.
I think that the 32-bit client should work on both 32- and 64-bit windows. If it doesn't we can try to compile a 64-bit version.
--> Ecco il perché delle WU molto corte:
In our simulations we need to perform work (saddle searches) for the current configuration of the chemical system in order to find what processes (events that change the configuration of the system such as an atom hopping on a surface) are available. Once we have done enough searches to meet our confidence goal we choose one of these processes and advance to a new state. Once we are in a new state no more saddle searches need to be done for the previous state. The work units are these saddle searches that we need to do some number of on average. This means that work units, in order to be useful, must be completed in about the amount of time it takes for us to complete all of these searches.
Currently for the systems we are studying and the potentials we are using a single saddle search takes less than a minute and we are making each work unit do 10-50 saddle searches. Each saddle search produces the same amount of data. Simply making each work unit take longer by doing more saddle searches will have two effects. First, the work unit results will be that much larger to upload (since the files grow linearly with the number of searches we do). Secondly, we will get our searches back in even bigger chunks. We might only need to do several thousand searches per state, which means for a given simulation (of which we run more than one at a time on BOINC) we never need more than several thousand searches being performed at once. This means that we would like to keep the number of saddle searches small in order to more efficiently parallelize the problem.
Another way of making work units take longer is to use more accurate potentials. The potential is the way the energy and forces in the chemical system are calculated. Right now we are using a very simple one which is why a single saddle searche take less than a minute. We hope to use more accurate ones soon, where one search might take many minutes or hours while still producing the same amount of output as our searches that currently take seconds. This would mean that one workunit would equal one search, which is very little data compared to the current work units and take much longer. These types of work units are better suited to distributed computing.