How many cpus are needed
Moderators: Global Moderator, Moderator
-
zhuanghl
- Newbie
- Posts: 31
- Joined: Thu Dec 08, 2005 2:15 pm
- License Nr.: 471
How many cpus are needed
Hi, I wanna calculate some system with more than 400 atoms using only 1 Kpoint. I would like to know how many cpus are generally needed for such a large system.Could someone who has similar experiences gives me some hint. Thanks.
Last edited by zhuanghl on Sun Apr 04, 2010 6:21 am, edited 1 time in total.
-
alex
- Hero Member
- Posts: 585
- Joined: Tue Nov 16, 2004 2:21 pm
- License Nr.: 5-67
- Location: Germany
How many cpus are needed
Hi, that would also depend strongly on the cut-off energy. But I'd guess that 128 cores should do the job in a some days.
Hth
alex
Hth
alex
Last edited by alex on Sun Apr 04, 2010 3:14 pm, edited 1 time in total.
-
Danny
- Full Member
- Posts: 201
- Joined: Thu Nov 02, 2006 4:35 pm
- License Nr.: 5-532
- Location: Ghent, Belgium
- Contact:
How many cpus are needed
That depends on a lot of things.
What type of cores are you talking about?
What you want to do?
What is your ENCUT ?
A 400 atom Pt/Ge slab with some adsorbad at ENCUT=400 eV and 4 IBZKPT did a static run in 46h on our local machine on 16 cores.
The relaxation was done on 64 cores on a supercomputer (Huygens) where it used a single kpoint and could do 1 ionic step in less than 1h (near the end)
You can just extrapolate from your current calculations.
multiply by the ratio of IBZKPTS
the volume ratio (V) gives an exponential time behavior (T=V^a, with a=1.9-2.2)
The efficiency of the supercomputer you wish to use is also something to check, normally the more nodes you use, the slower and less efficient it will be.
I would suggest do a test on a 40 atom system with the same cutoff, and extrapolate. It will not give you a 100% correct value, but it will give a good starting guess.
Danny
What type of cores are you talking about?
What you want to do?
What is your ENCUT ?
A 400 atom Pt/Ge slab with some adsorbad at ENCUT=400 eV and 4 IBZKPT did a static run in 46h on our local machine on 16 cores.
The relaxation was done on 64 cores on a supercomputer (Huygens) where it used a single kpoint and could do 1 ionic step in less than 1h (near the end)
You can just extrapolate from your current calculations.
multiply by the ratio of IBZKPTS
the volume ratio (V) gives an exponential time behavior (T=V^a, with a=1.9-2.2)
The efficiency of the supercomputer you wish to use is also something to check, normally the more nodes you use, the slower and less efficient it will be.
I would suggest do a test on a 40 atom system with the same cutoff, and extrapolate. It will not give you a 100% correct value, but it will give a good starting guess.
Danny
Last edited by Danny on Tue Apr 06, 2010 2:17 pm, edited 1 time in total.
-
panda
How many cpus are needed
Also, you could try doing a single point calculation first. Depending on how you compiled the program, what type of server you are running on, how the server is networked, etc. There is no one clear cut answer. If you post your computer specs and compiler information someone will be able to help you more. I have seen increases in speed up to 16 nodes, then decreases with anything beyond that on one system, then speed ups to eight nodes and decreases after that. It really does depend on your system and how things are configured.
Last edited by panda on Thu Apr 08, 2010 1:51 pm, edited 1 time in total.