Nicolas_orleans
[H]ard DCOTM May 2016
- Joined
- Oct 7, 2012
- Messages
- 352
Guys,
I know that the poor project management of FAH has driven most of us out of this DC project. However, there is some technical excitment coming in terms of bio computing. In case you haven't checked the thread in the beta section of FF, below the description that was made by Jesse_V.
Original post is here
Sounds cool in terms of science and tech, and I am pretty excited to finally see some AVX folding code being tested.
The fun part is : in less than one week BA is supposed to go 24 threads with reduced deadlines, and there was "surprinsingly" no communication or forward guidance, and probably another drama is on the way for the last BA folders. Will probably make more noise than this Ocores developement...
For the remaining beta testers
I know that the poor project management of FAH has driven most of us out of this DC project. However, there is some technical excitment coming in terms of bio computing. In case you haven't checked the thread in the beta section of FF, below the description that was made by Jesse_V.
Ocores are a really new type of core for F@h. It's not similar to FahCore_17 or other cores that you're familiar with. This is a completely new thing, with a new core, a new backend, and eventually a new front-end. You may have to escape the traditional simulation mentality to fully understand what's going on here. The reason that ocores are different is that they are streaming. Traditionally clients download a workunit, process it, and upload the result and then repeat, and each workunit has its own project-run-clone-generation number. Neither of these things are the case here. Now with this core, clients contribute to a simulation stream, their own protein trajectory. When a PG member wants to study a protein or another molecule, with the ocore architecture they first make what's called a "target". This is a base description of the protein. From the target the researcher can then make many "streams". Each stream relates to the target because it's the same protein; it has the same atoms, the same bonds, the same topology, but the difference is that each stream has a different initial set of atomic positions and a different set of initial velocities. For those that are familiar with the Simulation FAQ, a stream is basically a combination of runs and clones, a target is a project, and the generations are out the window. The new streaming design is not only more efficient, but it's supposed to be much easier to work with in the backend.
Another advantage of the streaming core is that deadlines are easier to handle and it's more efficient. Under traditional cores if I drop out of F@h then it takes weeks for my workunit to timeout and get reassigned. A stream core gains efficiency by synchronizing much more frequently. Now the results are returned to Stanford at every frame. This means that they get the results in a more continuous fashion. Also, every hour or so the Stanford servers are also sent "checkpoints", which are larger and more precise than frames but allow a stream to be resumed from that point. So now, if I go offline someone else can quickly pick up my stream where I left off, and there's less than an hour of downtime. Then I come back to the project I'd get a different stream to push forward. I also want to note that the PG researchers can stop and start a stream at any time.
At the moment, proteneer has created just one target, about 200 streams, and 10,000 steps per frame.
Original post is here
Sounds cool in terms of science and tech, and I am pretty excited to finally see some AVX folding code being tested.
The fun part is : in less than one week BA is supposed to go 24 threads with reduced deadlines, and there was "surprinsingly" no communication or forward guidance, and probably another drama is on the way for the last BA folders. Will probably make more noise than this Ocores developement...
For the remaining beta testers
Requirements:
-64-bit Ubuntu 12.04 LTS or later.
-Either a CPU with AVX support or an OpenCL compatible GPU.
-24/7 internet connection.
Download link: http://proteneer.com/siegetank/