Q6600 running SMP clients, 2 or 4 cores?

[Great_Gig]

I have a dedicated folding node with a Q6600 that runs Linux SMP 6.02 clients. I normally run 1 instance over all four cores, but was just wondering if it would be better to run 2 instances on 2 cores instead. Anyone know if one is better than the other?

Many thanks

[Ahavi]

Due to lagg between the dual cores running two SMP clients provide the best PPD, although is not the best for the project. There is a program for Windows that sets each client to use a specific pair of dual cores, but I dunno what your options are for Linux.

[Foxery]

The scaling problem is specific to Windows. You should only run 1 SMP client in Linux.

[v00d00]

Taskset can be used to set affinity.

Also it depends on whether you like points or science.

The choice is yours.

[Great_Gig]

Would rather have what's better for science, I'm way ahead in my team anyway, so PPD is not what's important. I want what's best for the project really.

So best for science is 1 SMP client over the 4 cores, is that correct? Don't really understand the comments about scaling problems or taskset being set to affinity

[Foxery]

The scaling problem is specific to Windows. You should only run 1 SMP client in Linux.

Don't really understand the comments about scaling problems or taskset being set to affinity

Er, I'll rephrase.

The reason for running 2 clients is specific to Windows. You should only run 1 SMP client in Linux.

[Great_Gig]

Thanks guys, 1 SMP client it is then.

[naapi]

Hmmmmmmmm,

I'm running 3 quads and on all of them running 2 clents of SMP yields better results (by 20% or more). I am able to still finish WUs within 24-36 hours so that's not delaying the program.

M.

[Ahavi]

Is that in Windows?

[naapi]

No, all quads run Linux - Ubuntu 7.10.

I understand 2 instances (2 clients) are unsupported by Stanford, but they still chyrn out much more, at least that's my experience.

M.

[Flathead74]

Hmmmmmmmm,

I'm running 3 quads and on all of them running 2 clents of SMP yields better results (by 20% or more). I am able to still finish WUs within 24-36 hours so that's not delaying the program.

M.

There is no way that slowing down the return of work units does also delay the project.

If you are seeing a gain of 20% by overbooking WUs, then there would also be a 20% slow down in returning the work.

I'm not trying to pick on you, but your statement just does not make sense.

[naapi]

I meant that if you can return the samples within the preferred deadline (2-3 days depending on project), the WU is not assigned to any other folder.

Therefore, by increasing PPD by 20+%, I'm in fact speeding up the projects, not slowing them down.

If, for example I get 6 WUs from the same project in a row, I am able to finish that 6 WUs within 5 days if I run 2 SMPs clients and within 6 days if I run just 1 client of SMP.

M.

[codysluder]

No, your logic is still faulty.

If you can complete one project in 1 day and you can complete two projects in 1.6 days then you are doing more work, but you are also slowing down the project. Every WU that you process wastes 14 hours when somebody else could have been working on the next WU in that trajectory. FAH is a lot more like a relay race than an event for a single person. The individual times of each leg of the race add up and four people can get better times than the average that a single person would do over the entire course. The next person needs to start processing the WU after 1 day, not after 1.6 days.

[toTOW]

Stanford recommends running one Fahcore per physical core ... so the recommendation for a quad core is to run only one SMP client ...

[Great_Gig]

Stanford recommends running one Fahcore per physical core ... so the recommendation for a quad core is to run only one SMP client ...

Thanks for that, that's what I have been doing and will continue to do so.