Rob, quick question, based upon the replies since I asked my previous question. From everything I'm hearing, wouldn't it just be better to scrap the massive mess, and try to write something fresh and new (and hopefully a lot simpler, and more scaleable/updateable), instead of trying to put a band-aid on a severed leg bleeding out of a femoral artery? Well, maybe not quite that extreme, as the patient probably has more than a few minutes to live, but you get the drift. Or, would the infamous DA be the impediment to doing such a thing? Face, egg, & all that such and stuff?

---

I've been looking at one for MB tasks, and it is actually quite trivial to do - a simple function of AR.

Function will be different for different devices.

Every time I look and see my rac dropping I keep thinking there's a problem then I remember Credit Screw

Not the cause anymore... it's now due to your exclusively NVidia farm receiving GUPPI VLAR work units on the GPUs. As noted much elsewhere, they process much more slowly than Arecibo MBs but pay the same credit.

I Wish it was just Nvidia being screwed but at one time I was over 20K for my rac with APs, without APs is was 12 - 15K. Right now I am struggling to keep it above 7K. I am running AMD R9 280. Not only is my RAC dropping due to the Guppis and new MBs, I have to shut boinc down when I do other stuff, I used to be able to both at once but now I get crashes if I do. Thats a double whammy on the old RAC.

Bob

---

Sometimes I wonder, what happened to all the people I gave directions to?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
SETI@home classic workunits 4,321
SETI@home classic CPU time 22,169 hours

You should get the SoG app. You will be able to run multiple wu's at a time for basically the same about of time it takes to run 1. Even without it, you can probably run multiple to better effect than you are now.

Chris

---

I assume I would download this file from rastimer

MB8_win_x86_SSE3_OpenCL_ATi_APU_r3430_SoG.7z

It looks the only ati sog file. I tried to find a set of instructions on how to install and run this but there doesnt seem to be a set of instructions I could find. I haven't used lunatics and from what I read it doesn't handle SOG with some tweaking.

While not a total luddite I am not sure I can handle a couple rewrite of my app_config.xml file without some help.

Thanx
Bob

---

Sometimes I wonder, what happened to all the people I gave directions to?
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
SETI@home classic workunits 4,321
SETI@home classic CPU time 22,169 hours

Al,
I wouldn't scrap the current mess until a suitable, maintainable, replacement was developed. The main reason is I wold want some of its data to confirm calibration and scaling.
I've just had a thought - given the issues I'm having getting enough suitable AP tasks I would set them to "constant score" for now, then the task-based score for MBs could go in quite rapidly.

As for impediments - "I couldn't possibly comment" ;-)

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

Once I get some free time at work, I'll find the link for you. You probably don't want the APU one, there is a better match for your card.

In the mean time, I would say grab the lunatics installer if you aren't opposed. It will help your CPU times and build your app_info.xml file so that all you will need to do is download the right SoG file and find/replace a few lines with app_info file.

Might be this evening but I'll send you a PM.

Chris

I assume I would download this file from rastimer

MB8_win_x86_SSE3_OpenCL_ATi_APU_r3430_SoG.7z

It looks the only ati sog file. I tried to find a set of instructions on how to install and run this but there doesnt seem to be a set of instructions I could find. I haven't used lunatics and from what I read it doesn't handle SOG with some tweaking.

While not a total luddite I am not sure I can handle a couple rewrite of my app_config.xml file without some help.

Thanx
Bob

---

For me, RAC will plummet to summer levels from now on, CreditNew or not. From this week, and until late August-September, I will only crunch from early evenings, to about 09:00 in the mornings.
It will get too hot in my apartment (a corner apartment with the sun shining right at it, from early morning, to early evening), and I refuse to run my portable AC just to be able to crunch. It isn't worth the kWh cost.

I'm prepared to hit the ground.

SPLAT!!

---

I'm just crunching for Milkyway for a week and already have a higher RAC, than i had here before... :O

Go figure! :þ

---

Aloha, Uli

I'm just crunching for Milkyway for a week and already have a higher RAC, than i had here before...

And if I move to Mexico and get a job as a janitor my paycheque in pesos will be larger than it is currently in dollars.

There won't ever be a BOINC consistent credit payment across projects, because smaller or startup projects or those with not-so-interesting (or nigh-incomprehensible) goals overpay to attract volunteers, and will continue to do so.

---

Since running GBT (vlar) MBs on my two machines' dual GTX 980s makes them run like dual 460s (or something like that), I decided to bite (buy?) the bullet and build a dual 8-core machine (there's a thread on it in Number Crunching) so I can have 30 or so HT threads running on CPUs; after I see how that goes for a while, I may delete my i7-4790K machine (too few cores) and move his 980s to the new guy.

Given that the vlars take about 4-5 times as long on the 980s as the non-vlars did, running 3 at a time on them will be the equivalent of maybe 3 cpu cores (it was 10 or 12 or more before the vlars). I may just sell the damn things, as they will have no advantage any more...well, except for APs, if we ever get any more of those. What a waste.

And that is because the CUDA app is miscoded - it works, but as we all know, runs very slowly on vlars. Perhaps some of our resident smart guys can fix it? I'm looking at you, Jason!!!!!

And this has nothing to do with Credit Ewww, which is the other bad code we have to deal with...

---

Perhaps some of our resident smart guys can fix it? I'm looking at you, Jason!!!!

I don't think it's fair to single him out. He already has a lot on his plate and donates his time and energy.

Everyone is in the same boat.

And if I move to Mexico and get a job as a janitor my paycheque in pesos will be larger than it is currently in dollars.

I knew this was coming, so i'm just the more pleased - thanks a lot! :þ

:))

---

Aloha, Uli

Perhaps some of our resident smart guys can fix it? I'm looking at you, Jason!!!!

I don't think it's fair to single him out. He already has a lot on his plate and donates his time and energy.

Everyone is in the same boat.

Oh I appreciate the 'faith', only warn that the time I have at the moment is pretty patchy/limited. Between all the apps being experimented with at the moment by others, and the knowledge we all have accumulated, I'm sure we'll find the better options. For now just assume it's going to take time, and there may not be simple one-size-fits-all solutions here. My contributions for the short term will probably be limited to testing some theories (with some special test pieces). For one thing it looks like speed/throughput on these is only a part of the problem, while scaling/load-balancing on vastly different machines may be a bigger part of the issues than before.

---

"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.

Down, down, and down it goes.
Where it stops, nobody knows.

---

Grant
Darwin NT

Ah well, it's quite clear now. With the introduction of GBT work, my RAC is falling like a stone on MB work. Since AP is just a "once per week" thing nowadays, my RAC doesn't have time to recover.

Thank you endlessly much CreditScrew.

. . Some things are funny, the Guppies MB tasks on my i5 run quicker than arecibo WU's and get a better ratio of credits/time, but on the GPU they take at least twice as long to run as Arecibo WU's and are only now beginning to show some proportionate increase in awarding credits, leaving them still way behind other tasks.

Yeah, the transient response on creditNew is pretty poor, and overshoot+oscillation is likely. [i.e untuned system]

---

"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.

IIRC we have the same conversation few years ago when V7 apears and the diference V7 vs AP comes to a big number.

At that time Jason and some others makes a lot of test (eben with other projects) and the conclusion was simple: creditscrew is broken.

It´s penalizes optimization.

Aparently DA and his team never agree of that and makes nothing to correct it.

Now the history is repeated.

My 0.02 cents.

Agree. What we need to do is sack DA, he is a millstone around ever ones neck ...

. . Time for a revolution brother ??? :)

. . Time for a revolution brother ??? :)

Revolution isn't the right word, we need a new one for 'Didn't realise what they made so now need to accept the future is not in their control.'

---

"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.

Perhaps some of our resident smart guys can fix it? I'm looking at you, Jason!!!!

I don't think it's fair to single him out. He already has a lot on his plate and donates his time and energy.

Everyone is in the same boat.

Oh I appreciate the 'faith', only warn that the time I have at the moment is pretty patchy/limited. Between all the apps being experimented with at the moment by others, and the knowledge we all have accumulated, I'm sure we'll find the better options. For now just assume it's going to take time, and there may not be simple one-size-fits-all solutions here. My contributions for the short term will probably be limited to testing some theories (with some special test pieces). For one thing it looks like speed/throughput on these is only a part of the problem, while scaling/load-balancing on vastly different machines may be a bigger part of the issues than before.

Hey, J, you can do it.

As I see it, you are correct, the problem with CN will not be solved easily. BUT that is not what is causing RACs to fall like a stone now. It is the fact that (given so many GPU MBs are vlars) the MB app has a bug in it that, while it still works, takes several times longer to complete for vlars than for non-vlars. Putting the effort into finding and fixing THAT problem will do a lot more than fixing CN. And should work across all machines, I think. So a more fruitful place to get credit "fixes".

SO: even if CN was giving "better" numbers, it can't do it for GPU vlars because, after all, they are really not several times the work, so the time consumed destroys the credit that machines like mine receive. Let me give 2 examples, based on (very) rough estimates of what my machines were/are doing.

One ("Machine A") is an i7-4790k running with HT off (for reasons of temperature control) running 2 threads of CPU. The other ("Machine B") is a Xeon E5-2670 running HT on, using 14 of 16 threads for CPU work. Both are running 2 x GTX 980 with 3 threads per card. So A has 8 threads total, and B has 20. Before vlars, on A, an MB on CPU took about 1 hour; on the GPU, 15 minutes. So A could do roughly 24x2 + 6x4x24 MBs/day, or 624. On B, it was ~1.5 hours on CPU and ~20 minutes on GPU, so B could do 18x14 + 3x6x24 = 684, so slightly more than A. And that's what I saw - A running a bit under 50K credit/day, B slightly over 50K/day.

But now, with almost entirely vlars on the 980s, which take about 4 times as long, we have the following situation: A -> 24x2 + (6x4x24)/4 = 192/day, B -> 18x14 + (3x6x24)/4 = 360/day. So even if CN was not a factor, Machine A would see his credit drop more than 2/3, and Machine B, almost 50%. Horrific.

That's the source of the current credit crash, and why I am going to move Machine A's pair of 980s to a new machine (see the thread about cheap 32 core machines) with 2 E5-2670s running with HT on, to get 30 CPU threads, so I can have some decent credit granted despite the bad MB app.

Anybody want to buy an i7-4790k and motherboard?

---