Why do I receive much less credit than my client claims although I use the latest client version 5.8.17? https://boinc.bakerlab.org/rosetta/results.php?userid=84658

It is a known fact that the Mac science app is only half as effective as the Windows client due to bad programming/optimization and therefore Macs receive only half the credit for the same amount of work done. Now I receive even less than the claimed credit (which is already much too low), due to what? Why do they even offer a Mac client when they do not want Mac-users to participate in this project?

---

it's not about not wanting mac users - most of the dev team use macs themselves. There are two factors involved. First, credit is falling generally so you aren't alone! My Athlons (Bartons) used to get much more credit than they claimed and now get around what they claim. There have been a number of threads on the subject, and there are probably a few causes, but I beleive the main one to be the more powerful cpus that have been released recently (i.e. conroe/kentsfield) which don't get such high benchmarks relative to the increased work they do. The reduced numbers of 'optimised' clients in the wild is probably also contributing.

Secondly, you're right that the client isn't ideally suited to ppc macs - I believe David Kim was looking at improving it, but then apple announced they'd be using intel x86 chips so investing dev time in the platform understanably became less of a priority.

Maybe an xbox client will increase the priority of the ppc architecture - I believe the xbox's xenon is similar to yours, but i could be wrong...

---

I had a very detailed review all prepared and got some error about the link timing out, and lost it all. But let me just throw this out there. I found a task that you crunched on your Mac, and I crunched on a Windows Intel. Please compare your task here, with mine here and let me know what you think about credit granted, claimed, fair, unfair whatever. By my calculations you received 50% more credit per second then I did, yet your benchmarks do not reflect that your machine does floating point that much faster then mine, and most of what Rosetta does is floating point.

[edit] and I would add that your machine appears to be able to crunch the 30 nstructs that mine did in about 52% of the time it took me to do it. So let's say your Mac is the benchmark system, all others are compared to it. You were granted 37 credits for your work. How many credits should I be granted based on your benchmark?

---

Add this signature to your EMail:
Running Microsoft's "System Idle Process" will never help cure cancer, AIDS nor Alzheimer's. But running Rosetta@home just might!
https://boinc.bakerlab.org/rosetta/

I had a very detailed review all prepared and got some error about the link timing out, and lost it all. But let me just throw this out there. I found a task that you crunched on your Mac, and I crunched on a Windows Intel. Please compare your task here, with mine here and let me know what you think about credit granted, claimed, fair, unfair whatever. By my calculations you received 50% more credit per second then I did, yet your benchmarks do not reflect that your machine does floating point that much faster then mine, and most of what Rosetta does is floating point.

[edit] and I would add that your machine appears to be able to crunch the 30 nstructs that mine did in about 52% of the time it took me to do it. So let's say your Mac is the benchmark system, all others are compared to it. You were granted 37 credits for your work. How many credits should I be granted based on your benchmark?

Feet1st,

this is not a matter of the client but of your specific setting. Your computer should be able to crunch a "standard" work-unit like the one I did in less than 9,000 seconds or the one you did in less than 30,000 seconds as opposed to the 85,000 you scored. Your settings may be responsible for this. You have to set the client so that it keeps the application in the memory and you have to set the "Switch between applications every" in the project preference page to a reasonable time (30-120 minutes).

---

Feet1st,

this is not a matter of the client but of your specific setting. Your computer should be able to crunch a "standard" work-unit like the one I did in less than 9,000 seconds or the one you did in less than 30,000 seconds as opposed to the 85,000 you scored. Your settings may be responsible for this. You have to set the client so that it keeps the application in the memory and you have to set the "Switch between applications every" in the project preference page to a reasonable time (30-120 minutes).

I'd be pretty surprised if 'the oracle' didn't have his settings right!

---

I think you are confusing RAC, with credit. If I did not leave applications in memory, and preempted to work on some other project (only other project I crunch is Ralph :) and was losing work, then my RAC would decline. However, my number of CPU seconds to complete the work I did would be nearly identical to, well indentical to the same machine configured in any other way.

If I've got 100min. in to a task, and then preempt it and force it out of memory, then upon restart I will see the CPU seconds drop back to the last checkpoint. Let's say back to 90min for example. So my total CPU recorded, and credits claimed would only be based on the 90min of work that survived. I will grant you that there are about 100 seconds expended in reinitializing a task after a full stop, and so I might report 100 seconds more then otherwise. But that's a 1% scale change. And so clearly not what we're talking about.

Are you confusing time on my wristwatch with CPU time? BOINC is reporting everything in actual CPU time. If my machine is configured to use 80% of CPU for example, you would never be able to see that in my credits for any specific work unit. This is because you would not know how many hours or days of time it took me to muster the 24hrs of CPU time reported. However, you WOULD see my RAC come down. But your question was not about RAC.

Yes, my settings are to crunch for 24hrs. That's my preference. And so we have to scale your information in to my runtime to account for that.

I think you probably have some other misconceptions about how credit works and I'd be glad to try and clarify things for you. Perhaps if you could explain how you concluded that my machine SHOULD have been able to do as many models in 30,000 seconds then I would better understand where you are coming from.

---

Add this signature to your EMail:
Running Microsoft's "System Idle Process" will never help cure cancer, AIDS nor Alzheimer's. But running Rosetta@home just might!
https://boinc.bakerlab.org/rosetta/

I think you are confusing RAC, with credit. If I did not leave applications in memory, and preempted to work on some other project (only other project I crunch is Ralph :) and was losing work, then my RAC would decline. However, my number of CPU seconds to complete the work I did would be nearly identical to, well indentical to the same machine configured in any other way.

If I've got 100min. in to a task, and then preempt it and force it out of memory, then upon restart I will see the CPU seconds drop back to the last checkpoint. Let's say back to 90min for example. So my total CPU recorded, and credits claimed would only be based on the 90min of work that survived. I will grant you that there are about 100 seconds expended in reinitializing a task after a full stop, and so I might report 100 seconds more then otherwise. But that's a 1% scale change. And so clearly not what we're talking about.

Are you confusing time on my wristwatch with CPU time? BOINC is reporting everything in actual CPU time. If my machine is configured to use 80% of CPU for example, you would never be able to see that in my credits for any specific work unit. This is because you would not know how many hours or days of time it took me to muster the 24hrs of CPU time reported. However, you WOULD see my RAC come down. But your question was not about RAC.

Yes, my settings are to crunch for 24hrs. That's my preference. And so we have to scale your information in to my runtime to account for that.

I think you probably have some other misconceptions about how credit works and I'd be glad to try and clarify things for you. Perhaps if you could explain how you concluded that my machine SHOULD have been able to do as many models in 30,000 seconds then I would better understand where you are coming from.

Feet1st,

if you compare your results to all the other computers with comparable CPUs you will learn what I mean. I have not found any other P4 running at 3 GHz that took so much time for that little amount of credit. It is as simple as that: Your computer receives less than half the amount of credit/time as any other Intel or AMD CPU of comparable speed. Probably several thousand users are wrong, but I tend to believe that YOUR settings are way off.

---

I didn't mean to imply that 1000s of users are wrong, nor that you are wrong, nor that I am correct. I'm just trying to explain what I know, learn what you think and see if there is an explaination for your observation.

If you could point me to a hostID or two that would support your assertion, I'd be more inclined to accept your conclusion. And in the big scheme of things, it averages out, but if possible it would be most pertenant if you could show me another host that crunched that same task that you and I did. That was what I was doing for you, to bring you truely comparable numbers.

---

Add this signature to your EMail:
Running Microsoft's "System Idle Process" will never help cure cancer, AIDS nor Alzheimer's. But running Rosetta@home just might!
https://boinc.bakerlab.org/rosetta/

I didn't mean to imply that 1000s of users are wrong, nor that you are wrong, nor that I am correct. I'm just trying to explain what I know, learn what you think and see if there is an explaination for your observation.

If you could point me to a hostID or two that would support your assertion, I'd be more inclined to accept your conclusion. And in the big scheme of things, it averages out, but if possible it would be most pertenant if you could show me another host that crunched that same task that you and I did. That was what I was doing for you, to bring you truely comparable numbers.

Feet1st,

your host requires 608 seconds for each credit (85750 divided by 141).
Here are a few comparable Pentium D hosts, some a little faster others a little slower than yours (taken from the Rosetta host stats page): https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=441048, https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=347756, https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=306640, https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=324688, https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=283508, https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=265888.

These hosts require 310, 421, 339, 318, 277, 429 seconds/credit. Pay special attention to the results of this host: https://boinc.bakerlab.org/rosetta/results.php?hostid=306640. It is a Genuine Intel(R) Pentium(R) D CPU 3.00GHz that takes appr. the same amount of time per work-unit but receives an average of 245 credits/WU (those are the big WUs).

---

Sorry, double post.

---

strange - i posted last night but it's not here...

Feet1st - it looks like your benchmarks are very low - my 1.33GHz P3 gets higher benchmarks than your 3GHz P4. (Mine's using the stock 5.8.16 client)

Are you running two jobs hyperthreaded? I would have thought HT would give the same benchmarks but just cause the time to increase more slowly...

---

I have a P4 3,0 HT this one https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=199440

It has almost the same bechmark as Feet1st.

Anders n

---

I have a P4 3,0 HT this one https://boinc.bakerlab.org/rosetta/show_host_detail.php?hostid=199440

It has almost the same bechmark as Feet1st.

Anders n

And my cpu concurs: P4 3,0 HT

HT computers recieve about half the benchmarks they deserve. I used to have a 3.2GHZ P4 (which i'm reviving soon) that had about 1,000/1,700 benchmarks. However, when I looked at other one-core processors of my speed range, they had way higher benchmarks. This is why it makes sense.

---

HT computers recieve about half the benchmarks they deserve. I used to have a 3.2GHZ P4 (which i'm reviving soon) that had about 1,000/1,700 benchmarks. However, when I looked at other one-core processors of my speed range, they had way higher benchmarks. This is why it makes sense.

The bechmarks is / cpu so mybe they they are ok after all?

Anders n

---

strange - i posted last night but it's not here...

Feet1st - it looks like your benchmarks are very low - my 1.33GHz P3 gets higher benchmarks than your 3GHz P4. (Mine's using the stock 5.8.16 client)

Are you running two jobs hyperthreaded? I would have thought HT would give the same benchmarks but just cause the time to increase more slowly...

From my experience with the Rosetta project and HT cores (I have a couple 2-core Xeons with HT so they each appear like 4 processors), allowing 4 "cpus" to crunch Rosetta results in a 10%-15% gain in RAC compared to only allowing 2 "cpus" to crunch. So apparently Rosetta bottlenecks on some other (shared) resource, probably the FPU. And no, it's not L1 and L2. I decided the small gain was not worth the extra heat generated. BTW, there is another project which I won't name were allowing 4 cpus comes close to doubling the RAC compared to 2 cpus. Different needs, etc.

---

dag
--Finding aliens is cool, but understanding the structure of proteins is useful.

Yes, I run HT. So, I guess what we've proved here is that Pentium D is faster then P4 HT. Yes, I leave tasks in memory. Yes, I crunch 24/7 and only crunch Rosetta and Ralph.

So, seems my benchmarks accurately reflect the work produced and the credit I receive.

Martin, rereading your original post, you stated that Mac receive half the credit for the "same amount of work done". Do you understand how Rosetta defines "work done"? Or did you mean to say for the same amount of runtime? Since all credit is issued based on work done, it would not be possible for anyone to receive half of what anyone else gets for the same work done.

Issues like preempting, and running multiple projects do not really effect your credit for a given task. They pertain more to how much of a 24hr day gets reported back as meaningful results, which impacts your RAC.

My machine benchmarks similar floating point ops per second to yours, with the same size L2 cache, yet you received more credit per second then I did. You have a Mac, I have an Intel. Different operating systems, etc. What I'm trying to get back to, was your original question. You seem to feel that Macs are singled out and treated unfairly. But I don't see it.

---

Add this signature to your EMail:
Running Microsoft's "System Idle Process" will never help cure cancer, AIDS nor Alzheimer's. But running Rosetta@home just might!
https://boinc.bakerlab.org/rosetta/

So apparently Rosetta bottlenecks on some other (shared) resource, probably the FPU.

Yes, HT CPUs still share a common resource for performing the actual floating point arithmetic. The other project you illude to must not be as floating point intensive as Rosetta.

---

Add this signature to your EMail:
Running Microsoft's "System Idle Process" will never help cure cancer, AIDS nor Alzheimer's. But running Rosetta@home just might!
https://boinc.bakerlab.org/rosetta/

Yes, I run HT. So, I guess what we've proved here is that Pentium D is faster then P4 HT. Yes, I leave tasks in memory. Yes, I crunch 24/7 and only crunch Rosetta and Ralph.

So, seems my benchmarks accurately reflect the work produced and the credit I receive.

Martin, rereading your original post, you stated that Mac receive half the credit for the "same amount of work done". Do you understand how Rosetta defines "work done"? Or did you mean to say for the same amount of runtime? Since all credit is issued based on work done, it would not be possible for anyone to receive half of what anyone else gets for the same work done.

Issues like preempting, and running multiple projects do not really effect your credit for a given task. They pertain more to how much of a 24hr day gets reported back as meaningful results, which impacts your RAC.

My machine benchmarks similar floating point ops per second to yours, with the same size L2 cache, yet you received more credit per second then I did. You have a Mac, I have an Intel. Different operating systems, etc. What I'm trying to get back to, was your original question. You seem to feel that Macs are singled out and treated unfairly. But I don't see it.

Feet1st,

in most other BOINC projects credit is based on the number of calculations done, regardless how much time a computer needs to do that. SETI and Einstein staff found that the calculation of credit based on benchmarks is B$ and only provokes cheating and therefore changed that. The problem is that the Mac science application for Rosetta is extremely inefficient due to bad programming. Therefore a PowerPC-based Mac needs twice as long to crunch a work unit than a comparable Intel or AMD processor based machine.

However, there must be something else going wrong here because most results I sent receive much less credit than the client claims and there is no other computer that did the same WU, so there is also no consensus like in other projects (3 results required, the middle claimed credit granted to all 3).

---

To use a quarum system wastes valueable CPU time. However 1000s of others crunched the same task, just not with the same seed starting point, and therefore there is some variation in the amount of CPU time it takes to complete any given model.

The credit you receive is based on the average of everyone else that crunched it. So average of thousands rather then midpoint of 3 redundant runs of exactly the same work.

To characterize a program that is not optimized for any given specific environment as "bad programming" is a very broad brushstroke you are making there. The topic of credit on Macs has been discussed before and I believe the net result of that was that compilers with the desired optimization capabilities are not available for the platform. Portions of Rosetta are coded in several programming languages, so you'd need several compilers all to have the desired optimization.

When someone ports code to a platform and uses the available compilers there and the code works, it's hard to say that the programming that was done was of poor quality.

---

Add this signature to your EMail:
Running Microsoft's "System Idle Process" will never help cure cancer, AIDS nor Alzheimer's. But running Rosetta@home just might!
https://boinc.bakerlab.org/rosetta/