Aah, lovely verbage... being American, I am of course, confounded by English.

One of my favourite phrases, brought to you courtesy of the United States Air Force:
The (mechanism in question) experienced catastrophic non-linear structural exasperation leading to energetic disassembly.
Use it carefully.

Cheers

---

Member of the 20 Year Club

One of my favourite phrases, brought to you courtesy of the United States Air Force:
The (mechanism in question) experienced catastrophic non-linear structural exasperation leading to energetic disassembly.

Lol, funny part is that makes absolute sense to me. Either the structural member in question was over-exerted, or the observer was a cunning linguist :-O

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

One of my favourite phrases, brought to you courtesy of the United States Air Force:
The (mechanism in question) experienced catastrophic non-linear structural exasperation leading to energetic disassembly.

Lol, funny part is that makes absolute sense to me. Either the structural member in question was over-exerted, or the observer was a cunning linguist :-O

Or both (8{)

---

Donald
Infernal Optimist / Submariner, retired

All this credit talk.. I figure this is a decent moment to pop in with a recent observation I've made.

So I'm still in the habit of keeping my spreadsheet of all the APs I've crunched, and aside from creditNew's obviously random nature per-task, I did notice something not-so-obvious.

So most of my APs take ~44,000 seconds to complete and depending on blanking, pulses found, interference from the Cosmic Microwave Background.. the typical credit ends up being between 650-730, with a strong bias toward the upper 600s. Every now and then, for reasons unknown to me, a task ends up stretching out to 48,000 seconds or even into the low 50s. Just one task does that..and the other two complete in a normal amount of time. After reporting, that task didn't even have high blanking.. in fact, often it usually has zero blanking..and also zero pulses found.

The interesting new observation is that this long-running task gets the typical credit granted..but then the next few end up being granted a bit less than expected (high 400s, low-mid 500s, typically). This only lasts for 2-3 validations and then goes back to the "normal" range.

I understand granted credit takes the recent average processing rate into consideration somehow, so it isn't unexpected to get less credit/second.. but the task that runs long gets normal credit and the next few get a bit less than they should. I assume that indicates the long task becomes the new "baseline" so to speak, and the next couple of tasks technically had less processing, therefore they should get less credits.

That's just another example of the oscillation that happens within one kind of task (not just when you switch from MB-only to AP-only). As far as I can tell, there isn't really much oscillation in RAC that I can discern. I did notice a long time ago that occasionally, if I get shorted on credits for a task (like say.. down in the 300s), some time within the next 10 tasks or so, I'll get one that grants 900+ to balance the overall average out.

Just my observations.

---

Linux laptop:
record uptime: 1511d 20h 19m (ended due to the power brick giving-up)

...That's just another example of the oscillation that happens within one kind of task (not just when you switch from MB-only to AP-only). As far as I can tell, there isn't really much oscillation in RAC that I can discern.

Yeah even filtering specifically to CPU MB shorties you see the oscillations. I believe the uncounted blanking plus the extra running time is enough to push it off balance. With MB the inherent variability of the tasks is one small contributor, though not that huge compared to the other issues. The bigger difference between MB and AP here is that stock CPU has AVX as well, so the perturbations are larger by giving the scale a harder shove (downwards) more frequently.

[Eric pointed out recently that the pfc_scales were going below 1, which theoretically is impossible unless applications use magic. Of course they don't use magic, but instead there are logic holes in the credit system that completely ignore SSE and AVX vectorisation]

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

there are logic holes in the credit system

Holes you could fit Olympus Mons through IMHO.
The abortion that is Credit New.

---

Grant
Darwin NT

there are logic holes in the credit system

Holes you could fit Olympus Mons through IMHO.
The abortion that is Credit New.

LoL. In some respects true (which is what makes it funny :P), but on the other hand the major issues are simple stability ones, so quite fixable without removing the intended benefits of self-scaling.

Having verified the overall high level design 'intent', as read from the code which does not match the whitepaper documentation, The things we're testing are lower level stability:

#1) coarse scaling error: caused primarily by omission of SIMD from the operation 'counts'. This causes drift, mostly downward, and AP-MB discrepancy. Fixable by taking SIMD and other future parallelism into account.
#2) oscillations (short term): caused by using sampled averages with no damping. That's basically the same problem DAC chips in cheap $5 CD players have. Fixing #1 then applying damping fixes this. The CD player equivalent modification would be the use of a much better DAC, with signal conditioning, resulting in a professional grade CD player.
#3) long term drift &/or slow response (too many tasks needed to respond to hardware change). Caused by averaging/filtering blurring out change. probably caused by using too many samples for averaging, which would have been done in a misguided attempt to address #2. After fixing #1 and #2, Add a delta term and fine-tune.

So yeah, it's a cluster, but fixable. Biggest challenge now is organising people across 4 continents, to get it all systematically tested for a well tuned/tested patch ( via albert@home , then probably beta here after). A bit like herding cats at the moment, and nothing to do with coding... just have to go with it.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Jason I am encouraged by your post, these people who keep slaming credit new with no solutions I find to be dispiriting. You and others are at least working on it. If mended Boinc will benefit and hopefully the madness with credit on some projects will cease.

---

So yeah, it's a cluster, but fixable. Biggest challenge now is organising people across 4 continents, to get it all systematically tested for a well tuned/tested patch ( via albert@home , then probably beta here after). A bit like herding cats at the moment, and nothing to do with coding... just have to go with it.

Good luck, with that & with new apps for Maxwell GPUs (not so subtle hint :-))

---

Grant
Darwin NT

Jason I am encouraged by your post, these people who keep slaming credit new with no solutions I find to be dispiriting. You and others are at least working on it. If mended Boinc will benefit and hopefully the madness with credit on some projects will cease.

Yeah, it's a bit of a case where the overall design intent (as read from code) is fine, just naive implementation. So the detractors are 'right' and so are the ones who don't lean that way... which makes for a good recipe for mayhem, lol.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

So yeah, it's a cluster, but fixable. Biggest challenge now is organising people across 4 continents, to get it all systematically tested for a well tuned/tested patch ( via albert@home , then probably beta here after). A bit like herding cats at the moment, and nothing to do with coding... just have to go with it.

Good luck, with that & with new apps for Maxwell GPUs (not so subtle hint :-))

Had some interesting insights there, and started some 'special tools' to help speed development. Basic idea revolves around building the infrastructure such that when the new full maxwell's come out, they optimise themselves. Mobile phone apps do it, so why not us ? lol.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Thank you Jason for telling us what is going on behind the scenes. My hat is off to you and the others who are fighting windmills:).

---

[/quote]

Old James

[quote] So the detractors are 'right' and so are the ones who don't lean that way... which makes for a good recipe for mayhem, lol.

This project in some ways reminds me of "The amature electrical hour" from Firesign Theatre from late 1960s and early 1970s

---

Thank you Jason for telling us what is going on behind the scenes. My hat is off to you and the others who are fighting windmills:).

+1 :)

---

Thank you Jason for telling us what is going on behind the scenes. My hat is off to you and the others who are fighting windmills:).

+1 :)

Any resemblance between myself and Don Quixote is entirely coincidental.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Thank you Jason for telling us what is going on behind the scenes. My hat is off to you and the others who are fighting windmills:).

+1 :)

Any resemblance between myself and Don Quixote is entirely coincidental.

Excellent Jason ! And who plays the role of Sancho Panza ? :p

---

Maybe Williams... LOL...

---

And 'El Rucio'.. :)

---

Noli tangere circulos meos...

Anyone know what happened to Bionc stats page? my current credit went from 2,993,272 down to 17,631 on Seti

---