8楼: Originally posted by souledge at 2012-05-25 22:47:27
OpenMP和OpenMPI是两个东西，第一个是内存共享的线程并行，第二个是进程并行MPI的一个实例化程序～前者在软件有代码支持的时候不需要外加程序就可以使用，在单机多核的并行效率比MPI高。后者MPI是无论是否编译时采 ...

11楼: Originally posted by hakuna at 2012-05-25 22:56:27
多敲了个I,那个地址应该是OPENMP,还有VASP测试结果...

NERSC provides a large number of precompiled Materials Science and Chemistry application codes, while majority of them are still flat MPI codes, eg., VASP, some of them have already had OpenMP directives added in the codes, eg., Quantum Espresso. To help to address the reduced per core memory on Hopper, and to take advantage of the imporved performance from using OpenMP, we chose two representative density functional theory (DFT) codes, VASP and Quantum Espresso codes, and analyzed their performance on Hopper upon the use of OpenMP and/or the multi-threaded BLAS library.

VASP is an application for performing ab-initio quantum-mechanical molecular dynamics (MD) calculations using pseudopotentials and a plane wave basis set. Currently it is the most frequently used DFT code at NERSC. VASP is written in Fortran90 and is parallelized with MPI. The current version of VASP (5.2.11) does not have OpenMP implemented in the code. However, by linking the code to the Cray multi-threaded scientific library (-lsci_mc12_mp) we are able to get some measure of the effects of threading upon its performance. This is a low-effort OpenMP implementation. We mesured the runtime and the memory usage of the two most commonly used iteration schemes, the RMM-DIIS and the blocked Davidson, for electronic iterations with two test cases that NERSC users provided.

12楼: Originally posted by souledge at 2012-05-25 23:04:02
就是这个地址，清晰的说VASP不支持OpenMP，所以用了其他的方法来测试并行效率。
首先文章介绍了测试的两个程序，VASP和PWscf，前者是用MPI，后者支持OpenMP：

然后是上面有人已经摘抄的一段：
...

13楼: Originally posted by hakuna at 2012-05-25 23:08:34
那这个怎么理解“However, by linking the code to the Cray multi-threaded scientific library (-lsci_mc12_mp) we are able to get some measure of the effects of threading upon its performance........”...

15楼: Originally posted by zouming1979 at 2012-05-26 00:00:47
用MPI写的程序都是send，recv之类的代码，用OpenMP主要是在循环的地方加上一些编译导语，指导编译器如何进行代码多线程化的编译。
上面的这句话的意思要看完整段话，大概应该是说部分代码直接可以链接到cray多线程 ...

3楼: Originally posted by dxf_dxf_dxf at 2012-05-24 19:20:06
如果你想用多个进程同时计算，就需要使用MPI。VASP是不支持OpenMP的。
如果你不安装MPI，一个计算任务就只有一个进程。
BTW，问一句CPU的个数和核的个数不一样啊？

14楼: Originally posted by zouming1979 at 2012-05-25 23:52:46
如果服务器的处理器是intel的，而且是几年前买的，肯定就是2个CPU，每个CPU4个核。
目前大多数我看到的vasp都是多进程版本，不是多线程版本的，因此如果不安装MPI，应该就不能跑并行程序，只能跑串行程序了。大多系 ...

2楼: Originally posted by hakuna at 2012-05-24 08:54:40
并行的话肯定需要MPI，并且按并行方式编译才可以....
如果不是必须并行，譬如做NEB或大体系计算或赶时间，那就无所谓了，你可以同时多加几个作业....

10楼: Originally posted by souledge at 2012-05-25 22:52:42
严格说，CPU个数是以sockets，也就是CPU接口为准的，说白了，就是买了一个CPU还是俩CPU。而核数是一个CPU内又细分的核心数量。
由于现在基本都是多核心的CPU，所以对于多socket的工作站和服务器，都是按照core *  ...