Page 3 of 3

Re: Not enough a7 WUs

Posted: Sat Mar 31, 2018 1:14 am
by DrBB1
Yup. And while a lot of folders are more tech savvy than average, many are not. And someone at the Lab must be a liaison with the tech people who develop/maintain the distributive software system. Assuming bruce's explanation is accurate, the techies and scientists could collaborate and adapt it for the blog. Is that idea really so unreasonable? Why, I'll bet even a group of biophysicists could do it! :ewink:

Re: Not enough a7 WUs

Posted: Sat Mar 31, 2018 5:11 pm
by bruce
You really shouldn't be expecting better collaboration between the scientists and the techies. They already collaborate effectively. It's not a question of flaws in the assignment logic. It's really a question of goals. :!:

The established goals of FAH are to maximize the science bei8ng produced. The science is project-based, not PPD-based. When a researcher launches a project, the simulation phase of the project is only a portion of the entire project. Thus there may be significant periods of time when there are only a few projects that need processing and there may of may not be other researchers who have projects in the simulation phase. The distribution of new projects depends on when some new projects finish the preliminary phases of the research and reach the simulation phase for that project.

FAH doesn't distribute Work Units tha can be called "busy work" (i.e.- repeating simulations that have already been completed.)

I recently attended a briefing at JPL where many space projects originate. One lab we saw was where many NASA research vehicles have been and will be prepared for launch. Although there are many projects in the planning phase, that particular facility was essentially deserted. Not long after that, I saw a story on the evening news that showed a new space vehicle being assembled in that same room.

Comparing JPL's lab with FAH's lab is not an ideal comparison. A lot of work has been done on NASA's Mars 2020 project whereas many FAH projects take a year or so. Nevertheless, it would have been fruitless for me to look at JPLs assembly room and ask why the FAHCore_a4 tools were being used more than the FAHCore_a7 tools, even if I had happened to visit when many activities were going on.

As I said earlier, the only active projects AT THIS MOMENT are limited to two servers at Stanford and two at Temple. There are 20+ FAH servers at 10+ universities within the FAH Consortium . That will change, but only when the scientists have prepared new projects for the simulation phase.

Re: Not enough a7 WUs

Posted: Sun Apr 01, 2018 1:32 am
by DrBB1
bruce, the collaboration I was referring to was for a blog post, nothing more. I assume the scientists and systems people already collaborate effectively or we wouldn't have F@H!

Re: Not enough a7 WUs

Posted: Sun Apr 01, 2018 3:05 am
by JimboPalmer
So far as i know, Mr Coffland is a one man shop. You want English out of him, but I want code. All the English comes out of designing/writing/testing code. I will miss his time.

I used to be the programmer for a gold mine, the auditors complained "Jim provides all our reports but never attends our meetings" I offered to attend rather than write reports for them, and they decided they liked the status quo. Mr Coffland is torn between Assignment Server code, Work Server code, Client code, Control code and (I think) Core code. Stretching him further will not help F@H.

Re: Not enough a7 WUs

Posted: Sun Apr 01, 2018 5:44 am
by Kuno
JimboPalmer wrote:You do realize that the folks at Pande Labs are Bio-Physicists, not computer techs, right?
With the amount of times the servers go down, and the sheer number of them that aren't even functioning..I think they've made it pretty clear they don't even know how to operate a computer, let alone be computer techs haha.

In all seriousness, they really need to start doing something to fix this tech issues of not using new hardware efficiently. It's like continuing to use an axe to cut down trees, when a laser saw is readily available.

Re: Not enough a7 WUs

Posted: Sun Apr 01, 2018 7:25 am
by JimboPalmer
Kuno wrote:With the amount of times the servers go down, and the sheer number of them that aren't even functioning..I think they've made it pretty clear they don't even know how to operate a computer, let alone be computer techs
I suspect they have so little budget, they are using some ancient hardware. By choosing the oldest to do the least important jobs, we see a great deal of stats server failure. (It is not vital to the science) A fair amount of 'server failure' is the University's network being made so secure, it fails to function. That is the various Universities network techs, not the Bio-Physicists.

Re: Not enough a7 WUs

Posted: Sun Apr 01, 2018 8:07 am
by JimboPalmer
Kuno wrote:they really need to start doing something to fix this tech issues of not using new hardware efficiently. It's like continuing to use an axe to cut down trees, when a laser saw is readily available.
For CPUs, a7 is pretty up to date. Once all existing research is completed, everyone with a CPU older than Haswell will howl that F@H does not love them. A7 performs significantly better on recent CPUs.

For Nvidia GPUs, F@H still supports OpenCL 1.2 from (2012+) Nvidia Kepler, Maxwell & Pascal GPU's (GeForce 600, 700, 800, 900 & 10-series, Quadro K-, M- & P-series, Tesla K-, M- & P-series) No newer OpenCL is supported by Nvidia, they prefer CUDA.

For AMD GPs, F@H could support OpenCL 2.0 (2011+) AMD GCN GPU's (HD 7700+/HD 8000/Rx 200/Rx 300/Rx 400/Rx 500-Series), some GCN 1st Gen only 1.2 with some Extensions
This would drop all the Terascale GPUs and some First Gen Graphics Core Next GPUs, say every thing designed before March 2013.

F@H would prefer to make one Core for both AMD and Nvidia, so staying on OpenCL 1.2 seems likely.

OpenMM has newer versions. (and OpenMM would be faster on CUDA, if F@H gave up on one Core for both AMD and Nvidia)
GROMACS has newer versions.

I suspect there are dependencies, if F@H stays on OpenCL 1.2, then I am not sure what version of OpenMM and GROMACS they can advance to.

Re: Not enough a7 WUs

Posted: Sun Apr 01, 2018 7:36 pm
by gbowman
Thanks all for bringing this to our attention. We've just released 4K new WUs and have more on the way.