Quad-core 2Ghz vs Dual-core 4Ghz - Which faster?

[shatteredsilicon]

Regardless, A2 scales very well, won't you agree?

No, I don't agree, because it is not reasonable to expect there to be no other load. Even if the machine is a dedicated folding machine, just running a GPU client will have a massive impact on the performance of the SMP client. Folding runs on spare rather than dedicated capacity in a vast majority of cases. The only way I can see the scaling issue being addressed is if the workload between the cores can be "unlinked", which I'm not even sure is possible given the nature of the work being done.

[shatteredsilicon]

Does Stanford say "Don't run this software unless you fold 24/7?"

As a matter of fact, yes, they do, in LARGE BOLD print on the High Performance download page. Not in those exact words, but the warnings and the VERY short deadlines easily translate to say run 24/7 or you won't make the deadlines, and so you won't get any points. That's a rather clear statement, run fast or no points.

Then they should then adjust deadlines to match that statement. The simple fact is that it only takes a 1.6GHz Core2 Solo to complete an SMP WU in time. Since this is about as low a spec 64-bit (i.e. SMP capable under Linux) CPU that can be had, is it reasonable to allow it to make the deadline even if it is running 24/7?

[7im]

Does Stanford say "Don't run this software unless you fold 24/7?"

As a matter of fact, yes, they do, in LARGE BOLD print on the High Performance download page. Not in those exact words, but the warnings and the VERY short deadlines easily translate to say run 24/7 or you won't make the deadlines, and so you won't get any points. That's a rather clear statement, run fast or no points.

Then they should then adjust deadlines to match that statement. The simple fact is that it only takes a 1.6GHz Core2 Solo to complete an SMP WU in time. Since this is about as low a spec 64-bit (i.e. SMP capable under Linux) CPU that can be had, is it reasonable to allow it to make the deadline even if it is running 24/7?

The deadline does already match that statement. It says if you have a dual core or quad core, including the slowest 1.6 GHz system, to the fastest stable overclock system, you should fold 24/7 to help the project the most.

Or did I mistake your post as support for excluding slower dual and quad hardware? Where would you draw the line and start excluding hardware? 2 GHz? 2.5 GHz? The SMP client already excludes some older Xeon (P4 based) quads, and now you want to exclude more?

[Sahkuhnder]

Then they should then adjust deadlines to match that statement.

In a perfect world the deadlines would only need to be long enough to barely allow time to finish a WU and then the results could be instantly and flawlessly sent back to Stanford. The internet we all love and hate is not up to that level yet. I'm sure you have seen the threads about server issues and problems sending completed WUs.

The deadlines allow extra time for the return of the results for practical considerations like power, network and server outages.

Just because the deadlines have a little extra time padding should not be interpreted to mean that Stanford does not prefer each WU to be completed and returned back absolutely as soon as possible. Their proposed new points scale reflects this fact as well.

The deadline lengths should also not be interpreted to mean that Stanford wants us to run the high-performance clients on older hardware that is barely capable of making the deadlines or that they want us to start doubling up SMP clients on a single CPU just because we can. The Pande Group has stated their policy on the optimal method for running their clients. I hope we can all agree that they know what is best and therefore all optimize our contributions by running the clients according to their very clear guidelines.

[shatteredsilicon]

Or did I mistake your post as support for excluding slower dual and quad hardware? Where would you draw the line and start excluding hardware? 2 GHz? 2.5 GHz? The SMP client already excludes some older Xeon (P4 based) quads, and now you want to exclude more?

Yes - I'm saying more should be excluded. You can't have your cake and eat it - either it's important enough to turn the results around faster or it isn't. The point here is that 1x 1.6GHz is enough to meet the deadline. Considering this is the SMP client, that seems quite ridiculous while claiming fast turn-around is so important. If it's an SMP client, then at least a dual core CPU should be required to meet the deadline. The slowest C2D is 1.8GHz, so if the recommendation is to fold 24/7, assuming a 75% duty cycle (about what you'll get out of it if it's not a dedicated folding machine), it should require at least a 2x 1.35GHz / 1x 2.7GHz machine to meet the deadline. According to that, the deadlines need shortening by about 40%.

[Ren02]

The slowest C2D is 1.8GHz

Core 2 Duo SU9300 runs at 1200MHz. There's even a 1066MHz part (Core 2 Duo ULV U7500) but that's nearly 2 years old.
Or should laptops be excluded from SMP folding? And all the AMD systems as well?

[shatteredsilicon]

I thought the official line is that "folding on laptops is not recommended" since most of them are sufficiently poorly designed that heat becomes an issue.

Anyway, it seems like a very conflicting set of requirements - fast turnaround and inclusion of low-spec CPUs.

[codysluder]

The question asked by the OP is fundamentally absurd.

If I have a 4 GHz machine, why would I want to run it at 2 GHz, no matter how many cores is has?

If you have a 4 GHz machine, it should be able to return WUs much faster than any deadline because it doesn't make sense to create a set of WUs with a deadline that can only be issued to 4 GHz machines. The whole idea is to create a project that can be done by a mix of 2 GHz machines and 4 GHz machines but which will have an average turn-around time that is better than assuming that every machine will only devote a fraction of it's resources to several projects based on the fact that the deadline is "too long" to demand using his entire machine.

The SMP issue is independent of that issue. If a dual core (or quad) machine can approach the speed of double (or quadruple) the speed of each of it's cores, why shouldn't Stanford take advantage of that speed by developing SMP software? At some point the development of that software will be completed, and until it is, the rest of this discussion is premature.

Yes, points should be based on speed, but is it fair to give the owner of a 4 GHz machine three times the points of the owner of a 2 GHz machine? There are already accusations that FAH is an elitist DC project, and that would confirm it. I firmly believe that those with old (slow) machines are still making important contributions to FAH, even if they can't complete with the speed of the newer machines. The whole point is that you need to contribute what you can contribute, not a fraction of what you can contribute.

[shatteredsilicon]

The question asked by the OP is fundamentally absurd.

If I have a 4 GHz machine, why would I want to run it at 2 GHz, no matter how many cores is has?

I think the main point in the original question was to establish if 1x4GHz was considerably better than 2x2GHz (which it is).

Yes, points should be based on speed, but is it fair to give the owner of a 4 GHz machine three times the points of the owner of a 2 GHz machine? There are already accusations that FAH is an elitist DC project, and that would confirm it.

Elitist vs. meritarian is a somewhat gray area here. If the speed of the response is important, then applying an inverse logarithmic valuation scale is the only sane way to make the points reflect the value of the contribution. Note: Linear scale wouldn't work, because that would not penalize running of multiple clients, and in fact, due to scaling problems mentioned earlier in the thread, multiple clients would still come out ahead in point production which is exactly what would need to be prevented.

[codysluder]

If the speed of the response is important, then applying an inverse logarithmic valuation scale is the only sane way to make the points reflect the value of the contribution. Note: Linear scale wouldn't work, because that would not penalize running of multiple clients, and in fact, due to scaling problems mentioned earlier in the thread, multiple clients would still come out ahead in point production which is exactly what would need to be prevented.

True, unless you can reliably measure actual GHz vs. effective GHz, and once you try to do that, it opens the client to a new type of cheating because "reliably" is impossible.

If Core A2 can scale within a few percent, the problem will mostly cease to exist.

[tear]

<snip>
(...) is it fair to give the owner of a 4 GHz machine three times the points of the owner of a 2 GHz machine?
<snip>

Are you saying this ^^^^^ currently applies to FAH? If so, can you provide some data please?
I'm just wondering how that's possible.


Cheers,
tear

[shatteredsilicon]

No, it doesn't currently apply to FAH, but it is being considered to stimulate low latency over high bandwidth of WUs.

[7im]

I think the main point in the original question was to establish if 1x4GHz was considerably better than 2x2GHz (which it is)

Did you mistype, or did I misread? 2x2GHz is better, for SMP client. 1x4 is better for CPU.

And we a bit passed the consideration stage for incentivising the quick return of work units. The SMP and GPU clients already do that. The next logical step is to do the same for all clients.

[WangFeiHong]

I thought we had established that 1x4Ghz processors were at least, faster than 2x2ghz processors due to inter-core bottlenecks.

[HaloJones]

Do people just not read posts or what.

Scenario 1.

3.2GHz Quad-core machine running Windows. Runs Win-SMP client with MPIEXEC 24/7. gets A2 unit and takes a whopping 26hours to fold it.

Scenario 2.

3.2GHz Quad-core machine running Windows. Runs two Linux VMs each assigned two cpus (maximum allowed under a VM). Each runs Linux SMP with mpiexec. each gets a2 unit and each takes 26 hours to fold it, resulting in the *same* folding time for each unit but does *two* in the same time.

Seems to me that I'm doing more for the project both in terms of return rates and total work done because I *work* at this.