I am using 6.6.20 on a 9800GTX, which is used half for SETI and half for GPUGRID.

I recently had a large download of SETI MB WU shorties into a refill of my cache - I had one in three being shorties, around 100 came in. They are of course the nature of the beast as a result of splitting, although such a large number at once is unusual at my end of life.

They all had short duration return, a few days, so in light of the number involved, I sorted the list, suspended the others, and I am now crunching through them, I should get through them by this evening.

I noticed that the credit given per shortie is around 30% less per cpu cycle than other MB WUs. I assume this is as a result of how the credit is granted compared to the need to load the WU into the GPU. It would seem on the face of it that when that aspect is brought into play, the resulting credit is about on par. However the real world end result is a credit imbalance per cpu cycle, however it is explained away.

Its not the end of Life As We Know It, as short WUs are relatively uncommon. This event has highlighted the imbalance because of the exceptionaly large number downloaded on this occasion. I doubt I would have noticed as such had it not been for the quantity in such a short time frame.

However over a longer time frame, such an imbalance can make a difference to the overall credit standing of the bigger Cruncher, who punches out MB WUs by the zillion daily compared to my small number. Not a huge difference, granted, but there non the less.

Would be good if it could be placed on the "things to do" list for the next Public BOINC (or application) release - whatever does the end credit calculation.

Regards
Zy

---

There have always been, in general, three "classes" of MB WUs. The "shorties" are (or were..due to the recent credit adjustments) typically ~15 credits, the "normal" ones (around 0.445 angle-range) came in at ~45 credits, and then there were long ones (high angle range?..I always get high/low mixed up) which varied, but were ~60-120. The long ones would deviate a bit regarding just how much the angle range actually was.

In all three scenarios, on long-tested CPU crunching (stock or opti), the credits/hour were vastly different. I can't really state a specific, or even approximate percentage difference betwixt the three, but basically..the shorties pay less, and the long ones pay more, per CPU-hour. Kind of like how r112 for AP_v5 pays more per hour than AK_v8 on the "normal" MBs.

From my understanding of CUDA and shorties, I think they end up taking..longer (?) than you would expect.


* Disclaimer: I have been awake for 22 hours now. The accuracy of my statements cannot be determined at this time.

[edit: alright, I got the highs and lows backwards. Shorties are high angles, the 60-120 credit ones are low angles.]

---

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

The CUDA shorties (and no doubt the CPU ones) do take up longer in terms of CPU cycles - and if in fact its a case of it all working out "even" with a slight over crediting (as such) on longer normal WUs, then thats fair enough, cant expect total precision to the Nth degree in all scenarios as there is an obvious pay off in terms of effort needed to code the differences compared to the long term utility / frequency of the coded event.

Would be nice to tweek this the next time the credit award code is revisited, if its easy enough to do without huge effort.

Either way no doubt the Earth will continue to spin on its customary axis without any sysmic shift due to change (or otherwise) in the code :)

Regards
Zy

---

The sticky bit is that for the regimes identified in the previous post, the ones that pay more and less are significantly influenced by architecture.

As the Conroe and Penryn architectures are quite common here, lots of people have noticed that the high angle range shorties tend to overpay for them. But I noticed on Mark's Nehalem machine that they underpay. They underpaid much more when it was running with seriously challenged RAM.

Now you plausibly state that for your particular graphics card implementation they underpay yet more.

While there are some inconsistencies across angle range in pay rate that are reasonably consistent across architecture, this is not one of them, and I think it is not subject to an easy fix.

---

Shorties, Oh yeah, I've got Loads of them here, Currently I'm running down My cache of 5.28 WU's so that I can go to 6.03, I know It's just a number, But Boinc 6.6.2x keeps saying that there's no work for My app and to Me that means 5.28 is old and 6.03 is current.

---

The T1 Trust, PRR T1 Class 4-4-4-4 #5550, 1 of America's First HST's

I get that and I'm running 6.6.20 with 6.08 CUDA. It started as soon as I went to 6.6.20, claiming none available for app type change options on website. However options are correct and not changed since running 6.5.0, so I can only assume its the difference on how the "new" schedular in 6.6.20 plays it. See how it pans out over the next few days, I've enough to keep me going on the card for five days.

Debt balance is zero'd - still the same result. I do run CPDN on the quad cpu, and I understand that can cause cpu priority issues, mine is in millions negative for CPDN (currently auto running at "high" priority). However its the only app I run on CPU.

See what happens in a few days, its possibly me not being used to the way the "new" schedular works and my mindset is wrong.

Regards
Zy

---

I just got through looking at the credits-per-hour of the various classes of units. I haven't done this for many months. It looks like the shorties give me about 30 percent more and the long ones about 20 percent more than the average ones, for both my PD950 and Q6600. By the way I like the improvement of the tasks page and the easily accessible stats and computer info with one exception: Every time I want to go thru the pages of uncompleted units I have to mouse the bottom of the page onto the screen so I can advance to the next page. It would be better to locate "Next 20" to the top portion of the page where I can click 1-2-3- etc without having to fiddle with the right panning facility for each page.

---

I just got through looking at the credits-per-hour of the various classes of units. I haven't done this for many months. It looks like the shorties give me about 30 percent more and the long ones about 20 percent more than the average ones, for both my PD950 and Q6600. By the way I like the improvement of the tasks page and the easily accessible stats and computer info with one exception: Every time I want to go thru the pages of uncompleted units I have to mouse the bottom of the page onto the screen so I can advance to the next page. It would be better to locate "Next 20" to the top portion of the page where I can click 1-2-3- etc without having to fiddle with the right panning facility for each page.

"Next 20" and "Previous 20" are both at the top and bottom of the page. I don't know why you have to go all the way down to the bottom to get "next 20", but it shows up at both ends of the page for me.

---

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

Thanks. I must've overlooked it. I'll check.

---