-bigadv benchmark machine

[Nathan_P]

Can I clarify what the current -bigadv requirement actually is?

- Is it a 16 'physical' core capability? or 16 'threads'?
- What is the memory requirement? - XXGB per core (or per thread) of a toal of YYGB?
- Do they run under both Windows and Linux?

I would say the requirements are:-

AMD: Dual 12 core 61xx running at 1.7Ghz or better, no chance of dual 8 core chips making the deadline 1.7Ghz might be cutting it close - i can't find any data
Intel: Forget any notion of 16 threads doing an 8101 in time, Forget any notion of 24 threads doing an 8101 in time if they are clocked lower than 2.66Ghz.

This basically puts most "server class" hardware not capable of -bigadv, Realistically to have any sort of longevity with regards to a machine doing -bigadv you are looking at quad amd on g34, Octo AMD on Socket F or quad intel on lga1567, not sure if quad hexes on socket 604 or socket F would make it??.

They should run under both but in reality you are needing *nix to run -bigadv at this time

[DocJonz]

Thanks;
- Glancing at Vijay's post, and the thread, tells me that its 16 physical cores. (bollix47 - your dual core Xeon system - is that 8c per CPU or 4c per CPU - as the former should work??)
- the history thread re: "extra large WU's" (is this what we now know as bigadv-16?) says 1GB per core is "more than enough" - 16GB for a 16 physical core machine should therefore be ample .....
- the thread mentions running on Linux (presumably x64), but what about Windows x64?

Perhaps I should cut to the chase - I'm toying with building a dual xeon E5 2670 machine and want to make sure that it can at least happily do bigadv-16!!

[Nathan_P]

Thanks;
- Glancing at Vijay's post, and the thread, tells me that its 16 physical cores. (bollix47 - your dual core Xeon system - is that 8c per CPU or 4c per CPU - as the former should work??)
- the history thread re: "extra large WU's" (is this what we now know as bigadv-16?) says 1GB per core is "more than enough" - 16GB for a 16 physical core machine should therefore be ample .....
- the thread mentions running on Linux (presumably x64), but what about Windows x64?

Perhaps I should cut to the chase - I'm toying with building a dual xeon E5 2670 machine and want to make sure that it can at least happily do bigadv-16!!

Bollix47's machine is 4c/8t per cpu

I've got 6gb in one of my -bigadv rigs and it coped fine with the wu, but 16gb does allow for future proofing. There is limited data for E5 26xx CPU's, i know of only one review that has tpf's so you will have to do some maths based on what clock speed you end up getting. http://www.servethehome.com/dual-intel-xeon-e52690-update-power-consumption-foldinghome/. It also has some reviews on dual lga2011 borads - but not the asus that i have my eye on

Win 64 won't work - the core is hitting 32bit memory limits when folding big WU, the linux cores are 64bit and thus don't suffer the same problem

[bollix47]

8c/16t = 8 cores/16 threads. Each of it's 2 CPUs have 4 cores/8 threads.

My 12c/24t is running @ stock of 2.66Ghz and takes 2.15 days on a P8101. So it is close as Nathan mentioned above.

[bruce]

My general rule of thumb is that when running Gromacs,Hyperthreaded (virtual) cores are worth about 70% of a real core. (It's only a very rough estimate.) The 4c/8t i7 was pretty close to [1 / 0.7] of the 4c4t i5. That makes bollix47's 24t machine pretty close to an equivalent of 18 real cores, thereby meeting the deadlines. The assignment server is allowed to distribute work to a machine with only 16 threads, but your goal is not to be able to GET the assignments, but to RETURN them. The real requirement is the deadline. There's no joy to be obtained by a machine with 8 real cores (actually FPUs) that is pretending to have 16.

[DocJonz]

Bruce - my aim is the opposite of what you've implied, i.e. I want to ensure that the 16 real core (plus the 16 accompanying hyperthreads) rig I proposed can give rapid turn round, and thus return, of the bigadv-16 WU's - I'm not in the game of 'just making it'!

[Nathan_P]

Bruce - my aim is the opposite of what you've implied, i.e. I want to ensure that the 16 real core (plus the 16 accompanying hyperthreads) rig I proposed can give rapid turn round, and thus return, of the bigadv-16 WU's - I'm not in the game of 'just making it'!

I would be looking at 2670 Cpu's as a minimum but in the UK they are not cheap, any slower and you might not have any wiggle room for bigger/slower WU. If you are looking for a dedicated f@h box then i would go for a quad AMD G34 machine, it will be cheaper and should last longer, for a daily driver then then xeon rig may be better

[DocJonz]

Quad AMD G34 sounds interesting in that case. I'm not up to speed on AMD CPU's (my last one was an FX-57!!) - what's the best options? Presumably these tend to be rack-mounted - but can you get towers to accomodate these?

[Macaholic]

Quad AMD G34 sounds interesting in that case. I'm not up to speed on AMD CPU's (my last one was an FX-57!!) - what's the best options? Presumably these tend to be rack-mounted - but can you get towers to accomodate these?

You can use a Xigmatek Elysium, or a HAF 932, among others, but they need some modifications. Being a team neutral folding help site, feel free to search the top 5 team forums, and I'm certain you will find several threads on the subject. Use keyword G34.

[bruce]

Bruce - my aim is the opposite of what you've implied, i.e. I want to ensure that the 16 real core (plus the 16 accompanying hyperthreads) rig I proposed can give rapid turn round, and thus return, of the bigadv-16 WU's - I'm not in the game of 'just making it'!

How can that possibly be? My aim was to quantify threads -- and provide some additional information about whether it's 16 threads or cores. Your aim is to quantify what it takes to get -bigadv. It sounds like our aims are pretty similar.

[DocJonz]

Bruce - my aim is the opposite of what you've implied, i.e. I want to ensure that the 16 real core (plus the 16 accompanying hyperthreads) rig I proposed can give rapid turn round, and thus return, of the bigadv-16 WU's - I'm not in the game of 'just making it'!

How can that possibly be? My aim was to quantify threads -- and provide some additional information about whether it's 16 threads or cores. Your aim is to quantify what it takes to get -bigadv. It sounds like our aims are pretty similar.

In that case, apologies, as I read your thread differently.

[7im]

The Assignment Server only gets a "CPU Count" from the operating system. In Windows, it reports threads as CPUs. So 8 cores/16 threaded processors could download the new -bigadv work units. But it wouldn't even come close to finishing the WU before the deadline.

And linux can be hacked to make is report more "CPUs" than it really has, regardless of actual cores or threads, so counting cores and threads means nothing at all.

The real concern is to return the WU before the deadline. If you can do it with less than 16 real cores, good luck, but you'll likely miss more deadlines than you make, and earn less points on average.

[DocJonz]

Perhaps in this case my posts have been mis-read or mis-understood - as the aim is that I want to ensure that the rig I build will be able to easily return all bigadv WU's well within any of the deadlines.

[Grandpa_01]

1 simpple answer without all the explinations MultiProcessor 4P amd 61XX 12 core CPU's on any SuperMicro H8QGi/6 and H8QGL boards due to the ability to OC them. These are currentley your best choice for bigadv, Plenty of time to spare

[7im]

And start with 2 chips, and you can add 2 more later.