Great idea! We would definitely increase our effeciency by a sizable amount. The question becomes, who writes the code, and how is it implemented?

Jean, I know we have posted on and discussed the subject of optimisation before, as have others, in many different ways, such as the recent thread on cache optimisations.
The difference here is that I am proposing a way for this to be taken on by the WCG centrally (Server side) or using the application (WCG application, or perhaps Boinc, using application or project update).
If you want to move this to better location, please do.
If I am shouted down, could someone else start a thread where we can all post our optimisation findings on our various systems. Then some smart guy with skills could produce a table of CPU by OS of task productivity, so everyone could benefit from the findings.
You wont get a badge for doing this but you will enable others to get more badges and do more work, and you will be able to use it yourself.

Jean, I have had one post deleted by someone else, one moved, and a tech even edited one of my posts by changing one word, but I have no idea how to delete a post.

I have checked through all the posts I made in the last month and the only one that is vaguely similar is, https://secure.worldcommunitygrid.org/forums/...ead,27970_offset,0#258054


You may be experiencing day-ja-vu!

Thanks for your reply Knreed,
I read the Boinc thread and understand Boinc is being adjusted WRT Credit for the better, and this will facilitate improved performance/productivity via system pairing and hopefully task allocation. All good stuff! I have made a few observations mainly WRT Credit towards the bottom.
I like your enthusiasm for this, but I am not quite sure how you will be taking account of the relative ability of the users system to crunch different tasks and allocate according to the general maximisation of productivity goal (your Matrix).
Will the system be able to spot the potential of a CPU that sits on a 64bit system to get 15% (not accurate) more work done when crunching HCC tasks than NRW tasks, for example? Or, will it, in effect, spot a CPU that is inherently best suited for crunching some tasks and allocate those tasks to it, because they would run much slower on other systems? I am thinking this may be the case, as the side effect would be more credit, but your approach seems to have a different basis (CPU-Pairing).


Credit - I know that you have to begin somewhere, and the scheduler’s estimate is just that, but I hope that this is at least regularly reviewed and adjusted. I expect the dynamic determination of CPU usages (performances) is the way forward with this, and would probably lighten background work in the long run and deal with the Computing Average drift following new batches of tasks, and perhaps new CPUs!

I agree that as with any science it is essential to identify anomalies and exclude them from corrupting the central data. So, it would be important to develop a system to adequately deal with exceptionally long and short results, even if it is just WRT Credit. You should however note that the actual data is potentially more important.
As a further aside, I can see a flaw with the GPU cross-project normalisation. Compute Capable factorisation needs to be included. 1.3 cards relatively outperform 1.1 cards by 41%. Obviously this requires better CUDA apps, and 3 is better than 2.3 and 2.1... but the CC 1.1 cards just cannot utilize this extra-efficient CUDA application coding! They will always be CC1.1. May I suggest you flip this and go by CC (card) rather than CUDA version?
You might want to think about the fact that G92 cards just cant handle some tasks. Unless the user aborts, the task will run until it fails and it will inevitably fail on that GPU. On other GPUs the task will run perfectly. This is not so much cherry picking as survival. People are leaving due to this.
PS. I like what you did here, http://www.worldcommunitygrid.org/forums/wcg/...ad,27770_offset,20#255617, and I can sort of see how you are trying to bring it all together
Good luck with it,