What about taking the granted credit of 100 WUs and calculate the credits per hour for each workunit and use these numbers to calculate an average?

Jochen

---

What about taking the granted credit of 100 WUs and calculate the credits per hour for each workunit and use these numbers to calculate an average?

Jochen

You can do that, that is sort of what Boinc does now, but what are you going to compare it too? The only thing that shows you is the trend on your own machine and as we have all pointed out, your machine cannot be compared to any other machine in any kind of meaningful way. Yes your 920 will be faster than my P3 or P4 but when trying to compare your machine to anything else there are just too many unknown variables. I have seen an old IBM pc XT overclocked from 4.77mhz to over 1ghz! It was the fastest XT on the planet, but could only do 1 thing, add numbers. And by that I mean 1 + 1 = 2 + 1 = 3, etc, etc ,etc. It was not comparable to any other XT at the time because it didn't do the same thing any other XT did!

I did the above for my 920 and 3 other 920s:

my 920: 17,7409880459578
Computer-Id 1143608: 17,4120945379865
Computer-Id 1141244: 17,0467817199461
Computer-Id 1159340: 18,0163583022697

These numbers are actually credits per hour per core. My 920 and the last 920 seem to oced to the same level, the other two do not seem to be oced (or they are using optmizted BOINC versions).
Looking at these numbers, it does not look like overclocking does make any difference. I still doubt very much, that it would make any difference, if I would rum my computer at stock clocks.
And just do not try to tell me, my computer is not overclocked correctly.

Jochen

---

For clarity, from the start:

Claimed Credit

This is based on the quick BOINC benchmark - 50% whetstone and 50% dhrystone, multiplied by the CPU time taken to process the Work Unit (WU). The benchmark is almost entirely a benchmark of the CPU core and doesn't take into account other factors such as cache size and memory speed etc.

Granted Credit

Each completed WU that your computer completes can contain many decoys (i.e. models). This is done to allow Rosetta to create work units that approximately
run for a desired time - 8 hours by default I believe(?). So one completed 8 hour WU might contain anything from 1 to 50+ decoys.

The granted credit is calculated as the average claimed credit from all previously submitted WUs of the same type, multiplied by the number of decoys.

An example

Take this WU: simpleF2_1f0s_2cx1_ProteinInterfaceDesign_15Apr2010_19616_136

My claimed credit was 47.97.
This was based on a benchmark that calculated I should get 15.98 credits per CPU-hour, and it ran for 10,806s (3.00 hours). 15.98 x 3.00 = 47.97 credits claimed. Therefore, if my benchmark had been twice as high, my claimed credit would have been twice as high.

My granted credit was 51.54.
The submitted WU contained 484 decoys. From previous submissions of this task, the R@H servers knew the claimed credit was, on average, 0.1065 credits per decoy. My computer therefore received 0.1065 x 484 = 51.54 credits.

So, in summary, if you computer gets a high or low benchmark compared to its R@H crunching ability, claimed will be high or low. Granted credit is based on the work completed so if your CPU has a large cache which doesn't help the benchmark but does help R@H crunching, you will probably find that claimed is lower than granted, and vice-versa.

The result is this:

PC 'A' has a small cache but a fast cpu core (e.g. 3GHz Dual-core AMD Athlon)
This computer gets a very high benchmark but can't match that on Rosetta.
On average it claims 50 credits per 8 hour task and is granted 40 credits.

PC 'B' has a large cache but slower FPU performance (e.g. 2.8GHz Dual-core Intel Core2)
This computer gets a lower benchmark but rosetta runs faster because of the additional cache.
On average it claims 30 credits per 8 hour task and is granted 40 credits.

One is claiming more than it receives and one is claiming less, but both are getting equal credit because their performance is equal on R@H.

HTH
Danny

---

I did the above for my 920 and 3 other 920s:

my 920: 17,7409880459578
Computer-Id 1143608: 17,4120945379865
Computer-Id 1141244: 17,0467817199461
Computer-Id 1159340: 18,0163583022697

These numbers are actually credits per hour per core. My 920 and the last 920 seem to oced to the same level, the other two do not seem to be oced (or they are using optmizted BOINC versions).
Looking at these numbers, it does not look like overclocking does make any difference. I still doubt very much, that it would make any difference, if I would rum my computer at stock clocks.
And just do not try to tell me, my computer is not overclocked correctly.

Jochen

I don't have Excel on this computer to work out the average credit per CPU-hour, but if yours is lower could it be because they're using Vista (which I believe is more efficient at scheduling more cores, and you have lots), or because they're using x64? I'm not sure what difference this would make but it might be one of those, or possibly that Rosetta is getting a boost from turboboost on the stock machines which I presume is disabled on yours due to the overclock?

---

I wrote a little program in C++ for this. I do not trust Excel... ;) No, actually it took me less time to write a little parser than preparing the data for an Excel sheet.

And of course this is different to the BOINC-RAC, since it does not take into consideration in what period of time these results were provided. With this credits per hour per core you should be able to calculate the RAC quite exact, if your computer runs 24/7.
The only reason for divergences is whether or not the is load from other programs _while_ crunching.

Jochen

---

I have now parsed some data from my other computer (Q9650).

Credits per hour per core: 21,0304418341766
21,0304418341766 * 4 (cores) * 24 (hours) = 2018,9224160809536 RAC. This is pretty close to the real RAC of currently 2005: https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=1096431

Jochen

---

If you happen to be curious how BOINC RAC is actually calculated. This is a link to a wiki page about RAC:
http://www.boinc-wiki.info/RAC

---

For clarity, from the start: ...

Thanks for that. That's the neatest explanation I've seen.

---

Oops - it should say:

The granted credit is calculated as the average claimed credit per decoy from all previously submitted WUs of the same type, multiplied by the number of decoys.

:D

So what you are trying to tell me, is that I got a fast machine, but Rosetta can not make any use of it?!? In this case, I rather quit crunching Rosetta on this machine...

Jochen

---

So what you are trying to tell me, is that I got a fast machine, but Rosetta can not make any use of it?!? In this case, I rather quit crunching Rosetta on this machine...

Jochen

No, the 920 is about as good as you can get on rosetta, and an overclocked one is even better. What i'm trying to tell you is that a machine that consistently gets less credits than claimed can still be getting more credits per WU than a machine that gets more credits than claimed. A difference between claimed and granted is irelavant because claimed is based on a benchmark that doesn't reflect rosetta very well and so can be artificially high or low.
To put it another way, if we both ran exactly the same WU for 8hrs and I was using an old athlon we'd get the same credit for each decoy processed but yours would complete many more decoys in 8hrs than mine would so you'd get proportionally more credit for that WU. I'll have a look at your numbers this evening and get back to you though.

---

So what you are trying to tell me, is that I got a fast machine, but Rosetta can not make any use of it?!? In this case, I rather quit crunching Rosetta on this machine...

Jochen

No, the 920 is about as good as you can get on rosetta, and an overclocked one is even better. What i'm trying to tell you is that a machine that consistently gets less credits than claimed can still be getting more credits per WU than a machine that gets more credits than claimed. A difference between claimed and granted is irelavant because claimed is based on a benchmark that doesn't reflect rosetta very well and so can be artificially high or low.
To put it another way, if we both ran exactly the same WU for 8hrs and I was using an old athlon we'd get the same credit for each decoy processed but yours would complete many more decoys in 8hrs than mine would so you'd get proportionally more credit for that WU. I'll have a look at your numbers this evening and get back to you though.

Jochen has a 3 to 4 day cache so you will have to go a few pages to get to some completed units, but he is farily consistently doing units in the 1 to 2 hour range and only doing about 5 to 10 decoys per unit. At least he is returning units in that time frame, 3 to 4 days, and you need to go all the way to 240 workunits to find units that have been returned.

Jochen has a 3 to 4 day cache so you will have to go a few pages to get to some completed units, but he is farily consistently doing units in the 1 to 2 hour range and only doing about 5 to 10 decoys per unit. At least he is returning units in that time frame, 3 to 4 days, and you need to go all the way to 240 workunits to find units that have been returned.

Basicly correct. It is a 3 day cache, but it might be messed up a bit, since my internet was 'broken' for 24 hours until 10 AM today (MESZ = GMT+2). Currently the first page with done results start at 220 (hit the Next-button once on the first result page and change the end of the URL from 'offset=20' to 'offset=220').

From my calculations, my 920 shuold reach a RAC of 3500. I was rather expecting a RAC close to 4000...

Jochen

---

I have just processed some more data form my 920:

Processd results: 177
Min. credits per hour: 13,7128212108908
Max. credits per hour: 26,3064721433467
Ave. credits per hour: 18,2945678916994
Estimated RAC: 3512,55703520629

How is it possible, that the credits per hour vary bei 50 percent? I was not running anything else but Rosetta on this computer. Is this a hyperthreading issue?

Jochen

---

I have just processed some more data form my 920:

Processd results: 177
Min. credits per hour: 13,7128212108908
Max. credits per hour: 26,3064721433467
Ave. credits per hour: 18,2945678916994
Estimated RAC: 3512,55703520629

How is it possible, that the credits per hour vary bei 50 percent? I was not running anything else but Rosetta on this computer. Is this a hyperthreading issue?

Jochen

I just realized your PC has 3GB of RAM... and is running 8 threads of JUST Rosetta... That's even below the 512/core recommendation.

Not sure whether HT hinders or helps out Rosetta crunching.

---

I just realized your PC has 3GB of RAM... and is running 8 threads of JUST Rosetta... That's even below the 512/core recommendation.

I am aware of this problem. Actually it has 6 GB of RAM, it is just the limiting 32-bit OS... As well I have been watching the memory consumption since I reinstalled BOINC last Saturday. Average usage is 300 MB per WU, with a peak here and there up to 320 MB. But usually there are still 400 MB of RAM left...

Should not be much of a problem. I did not see WUs 'waiting for memory'.

Jochen

---

Not sure whether HT hinders or helps out Rosetta crunching.

From previous posts it helps on i7 - not so sure on P4.

---

The granted credit is calculated as the average claimed credit per decoy from all previously submitted WUs of the same type, multiplied by the number of decoys.

That is even worse than I expected. What is it good for granting credits by the average claimed credit for a decoy?!? There are so many different computers out there that this does not seem to be a fair method at all. As well using an optimized BOINC client will result in claiming far to much credits - as long as there is no optimized Rosetta client.

I would run Rosetta even without any crediting system... But since there is one, it should be fair and reasonable.

There sure is a way to count the actual FLOPS for a WU. In my mind this would be a better base for a crediting system.

Jochen

---

The granted credit is calculated as the average claimed credit per decoy from all previously submitted WUs of the same type, multiplied by the number of decoys.

That is even worse than I expected. What is it good for granting credits by the average claimed credit for a decoy?!? There are so many different computers out there that this does not seem to be a fair method at all. As well using an optimized BOINC client will result in claiming far to much credits - as long as there is no optimized Rosetta client.

I would run Rosetta even without any crediting system... But since there is one, it should be fair and reasonable.

There sure is a way to count the actual FLOPS for a WU. In my mind this would be a better base for a crediting system.

Jochen

It's fair - I think you misunderstand?

I'll give another example:

If my slow computer finishes a decoy in 10 mins and your computer (which is 5x faster) completes the same decoy in 2 mins then we'll both receive the same amount of credit for that decoy (because they've done the same amount of work). In a given period, your computer will complete 5x as many decoys as my computer and so will get 5x the credit. Likewise, a 30% overclock should result in 30% quicker completion of each decoy, and so 30% more credit in a given period.

The claimed credit has very little effect because the value of each decoy is calculated from the average credit claimed for all previous decoys of that type. Therefore if you have a massively high benchmark (and consequently an excessive claimed credit), it will have a very minor effect on the granted credit - it will only serve to slightly increase the average claim for that type of decoy.

Danny