Cores a3 and a4

[iceman1992]

I noticed that I never get a3 work units, always a4. I'm just curious why that is. Do I need to add some flags like max packet size big or something?
And why do a3 projects have much higher k factors?

[Joe_H]

Partly it depends on which client you are using and the settings such as normal vs. advanced. Then it is up the the assignment servers, they will assign your machine to work servers based on availability and need.

As for the k-factor, A3 WU's have short deadlines compared to similar sized A4 ones. A higher factor is needed to give a high enough QRB. A4 WU's are a hybrid of longer deadlines so as to be suitable for uniprocessor machines, and being able to be done on a SMP setup.

[iceman1992]

Partly it depends on which client you are using and the settings such as normal vs. advanced. Then it is up the the assignment servers, they will assign your machine to work servers based on availability and need.

I am using normal I think. I added no flags. Default from the installation. So if I want to get some a3 WUs, what flags do I have to add?

As for the k-factor, A3 WU's have short deadlines compared to similar sized A4 ones. A higher factor is needed to give a high enough QRB. A4 WU's are a hybrid of longer deadlines so as to be suitable for uniprocessor machines, and being able to be done on a SMP setup.

Do they take longer to complete than a4 WUs?

[Nathan_P]

i've found that my v6 clients usually get a3 WU on either linux or windows, my v7 windows client has run nothing but a4 WU since i installed it 6 weeks ago.

Not always, there are short and long WU for both cores - you can't compare exactly though as each project will normally only be assigned to a particular core, this makes comparisons difficult

[iceman1992]

i've found that my v6 clients usually get a3 WU on either linux or windows, my v7 windows client has run nothing but a4 WU since i installed it 6 weeks ago.

I don't remember when I was running v6. Now my v7 has been getting a4 all the time.

Not always, there are short and long WU for both cores - you can't compare exactly though as each project will normally only be assigned to a particular core, this makes comparisons difficult

Yes well just a rough comparison, on average. No need for accuracy, just for curiosity's sake

[Joe_H]

Part may be connected to the assignment servers connected to by the clients. With V7 "assign3" and "assign4" are the ones checked first by the client. Other versions of the client appear to default to assign.stanford.edu from what I see in old logs. The A4 core also needs a minimum version of the client to be downloaded and used, so if you ran an older version of V6 with SMP you would just get A3 WU's.

OS also matters a bit. For example, there is a group of A3 projects that are currently only assigned to OS X and the V6 client. They use the older v2.22 A3 core, a later version of the core is available only for Linux and Windows. Now that I am using the V7 client, it has been months since I have processed one.

Do they take longer to complete than a4 WUs?

Depends on the size of the WU's. I did mention that for comparing ones of similar size the A3 deadlines were shorter. As an example, here are the figures for two projects from the summary that I have done in the past and took about 24-26 hours to complete (similar size) on my iMac.

Proj:6040 atoms:330706 pref-dl:6.00 final:8.00 base:1395.00 GRO-A3 k:4.63
Proj:7808 atoms:59392 pref-dl:25.23 final:54.66 base:1698.09 GRO-A4 k:0.58

The comparison is not exact, the 6040 took just over 24 hours, while the 7808 was between 25 and 26. The A4 deadlines are much longer. Since the QRB formula starts with the ratio of processing time to deadline, if the k-factor were the same the bonus would be much too large for the 7808 or too small for the 6040.

[bruce]

Depends on the size of the WU's. I did mention that for comparing ones of similar size the A3 deadlines were shorter. As an example, here are the figures for two projects from the summary that I have done in the past and took about 24-26 hours to complete (similar size) on my iMac.

Proj:6040 atoms:330706 pref-dl:6.00 final:8.00 base:1395.00 GRO-A3 k:4.63
Proj:7808 atoms:59392 pref-dl:25.23 final:54.66 base:1698.09 GRO-A4 k:0.58

The comparison is not exact, the 6040 took just over 24 hours, while the 7808 was between 25 and 26. The A4 deadlines are much longer. Since the QRB formula starts with the ratio of processing time to deadline, if the k-factor were the same the bonus would be much too large for the 7808 or too small for the 6040.

WUs depend both on the number of atoms AND on the number of time-steps. plus other significant factors that are less visible. A protein with twice as many atoms potentially has 4 times as many inter-atomic forces to be calculated though there are some which can be neglected. If it takes longer to calculate the forces in a protein, they may choose to create each WU with fewer steps, somewhat compensating for what we commonly call "bigger" proteins and keeping the processing time within a reasonable range.

Points Per WU is generally a better indication of expected processing time than JUST atom counts.

[iceman1992]

So effectively a3 and a4 gives similar total points, and thus similar scientific work? I read a post saying that a3 gives much higher points some time ago. That's not true then?
If a4s had much longer deadlines, why did P8013 have a pref-dl:3.32 and final:7.20?

[bruce]

, , , why did P8013 have a pref-dl:3.32 and final:7.20?

You're a Stanford scientist with four projects. Two are high priority; two are lower priority. Two are computationally difficult; two are less complex.

The high priority tasks should be assigned only to (say) quads or above. Unprocessors can be helpful on the lower priority tasks even if they're computationally difficult but probably the simpler tasks should be preferentially assigned to uniprocessor clients. Shorter deadlines can be assigned to the SMP projects and longer ones must be available to the uniprocessor WUs but within each category, the value depends on the computational complexity as well as other factors.

Managing projects at the Assignment Server level is a complex task.

[iceman1992]

So does that mean a3's deadlines aren't much shorter than a4's? And if I spot a project that is assigned to both SMP and uniprocessor, then does that mean the project is not high priority?

[7im]

a3 deadlines (as a ratio to their size) are shorter in general. The larger K factor shows that. However, over all length could be longer or shorter than a4s.

While that may seem logical, you cannot assume a4 projects are low priority. a4s are going out to both CPU and SMP clients, so a much larger number of clients are working on these projects, which may help get them all completed faster instead of slower. The number of runs clones gens has an affect on which way the clients would finish faster. If a narrow band of projects, with a need for many many generations, then SMP only would be better. If a wider band of projects, with fewer gens needed, then running on both CPU and SMP would help.

[iceman1992]

While that may seem logical, you cannot assume a4 projects are low priority. a4s are going out to both CPU and SMP clients, so a much larger number of clients are working on these projects, which may help get them all completed faster instead of slower. The number of runs clones gens has an affect on which way the clients would finish faster. If a narrow band of projects, with a need for many many generations, then SMP only would be better. If a wider band of projects, with fewer gens needed, then running on both CPU and SMP would help.

I see. Well that explains it. So the conclusion is that neither a3 nor a4 are of a higher priority than the other?

P.S. I still don't know what runs, clones, and gens mean
Edit by Mod: Posts on that topic have been moved to a new thread: viewtopic.php?f=16&t=21616

Btw thank you Joe_H, Nathan_P, bruce, and 7im you have all been very informative

[bruce]

Managing projects at the Assignment Server level is a complex task.

7im pointed out an important factor that I forgot to mention: How is a particular project structured in terms of serialization/parallelization? Most FAH projects require a large number of Gens and each one is generated when the previous Gen is completed. The total time to simulate a single trajectory is greater than the sum of the times it takes to run each Gen. Thus those projects benefit most from running on hardware with lots of cores. Some projects only need a few Gens but can benefit from a larger quantity of Runs or Clones. Assuming equal computational complexity in a WU, it's less important whether a WU runs on a fast or a slow machine. Lots of uniprocessors can work on the same project in parallel and turn-around time for each Gen is less important even though the project might be just as important.



Projects like the latter are less common in FAH but when they have them, there are lots of WUs that can be assigned to ANY computer. Projects like Seti@home are the ultimate extreme for easy parallelization. Any segment of the sky can be processed by person A concurrently with any other segment by person B. Speed of individual segments are not important . . . just the percentage of segments that have been completed.