Discussion Thread for the Continuous GW Search known as O2MD1 (now O2MDF - GPUs only)

again its not about "listening" correctly. it's about ASKING correctly in the first place. and the query being used in petris  code is not the same as either the query being used in BOINC, nor the fix the Ville implemented. BOINC right now says "What is the available memory, and give me the answer as a 32-bit number", Ville's code says "What is the available memory, and give me the answer as a 64-bit number", this is generalized from the different API calls being implemented between the two.

--edit---

this is what the Developer says:

reading a little more about the limit, it seems like it's still there (as I can find recent posts about it), but the limit is only implemented for a SINGLE buffer. I'm sure most science apps are getting around the limit simply by loading up multiple smaller buffers using less memory each. which makes sense for a highly parallelized application running on a processor with thousands of cores available.

it's just not a factor in why the tasks are failing here on cards with 3GB or less.

_________________________________________________________________________

I think we're talking cross issues here. I initially only questioned why you mentioned what BOINC detects, as it has nothing to do with the OpenCL 27% limit. BOINC incorrectly getting memory size on nvidia has nothing to do with the openCL limit for single buffers.

you said

and I merely commented that 'BOINC', while getting the wrong value, has nothing to do with it. If it did, then the incorrect value it pulls might limit you to only being able to use 4GB, but it doesn't so no point mentioning what BOINC does or does not see. its all up to the science app at that point. but there is no problem there. 

_________________________________________________________________________

I'm not saying that it is. I'm explicitly saying that's not the case. I don't know how you could interpret it any differently. You're the one that brought up BOINC's total available memory detection. But the fact is that whatever BOINC detects has nothing to do with what the science app is doing, they are separate entities. anything printed in the stderr file is done so by the Science App, NOT by BOINC. it's a simple concept.

as far as I see it, if you try to use a single buffer in your code, then you can't use more than 25% of the memory, BUT if you instead parallelize with multiple buffers running simultaneously, you can use 100% of your memory, provided that each single buffer is under the limit. this is where the 3GB problem is coming into play most likely at Einstein. since the sum of all the individual buffers is exceeding 3GB Total available on the card. nothing to do with detection, everything to do with just running out of space. 

_________________________________________________________________________

I never said that it didn't. you seem stuck on arguing a point that no one is making. the fact that BOINC reports correctly for less than 4GB is immaterial to the issue at hand. It literally has nothing to do with anything, except to be used as a diversion tactic I guess?

the point I made was that whatever BOINC reports in NO WAY affects what the science app is doing. BOINC could report 0GB of GPU ram and the science app would still work just fine.  YOU mentioned the numbers that BOINC reported first. I said they don't matter, and you proceeded to deflect and argue with me until you came full circle to claim they don't matter to the science app, which is exactly what I said in the first place LOL, can't make this stuff up. I only brought up the issue of BOINC reporting the incorrect VRAM for nvidia cards as a foot note to illustrate my knowledge of the issue since you seemed to be implying every app is getting the data the same way and that somehow BOINC just "listens wrong", which is objectively incorrect. there are multiple methods that can be used. fixing BOINC's issue is a nice to have, not a necessity, which also highlighted the point that whatever BOINC detects doesnt matter in actual processing (because if it did, then people running 2x with GW on <6GB cards would also see failures, but they dont. so this must also be incorrect).

the developers CAN'T correct the 27% issue even if they wanted to, nvidia themselves seem to have control of this. it's not so much a myth as it seems to be misunderstood. with people interpreting it to be a limit on total use when the limit is actually only applied to single buffer use. you get around the single buffer limit, by using multiple smaller buffers, allowing you to use all of the available memory if you wish.

_________________________________________________________________________