Query about using shared memory

zako

Hi everybody,
It is possible to use the memory of multiples nodes with the CPU of one node in calculation of gaussian 09, in purpose to gain some speed, instead of using TCP-Linda which is very expansive.


Sincerely,

liyuanhe211

I'm not sure what you mean by "use memory of multiple nodes".

If you mean that you connect the two nodes in some way, like infiniband, then perform cross-node computation in some way (other than Linda?) using only CPU(s) in one of the nodes. I couldn't see any advantage of doing this. Seems to me it might even be slower than just using one node. If you can do cross-node computation efficiently, why not just use all the CPUs? Also I'm not sure there is a way other than Linda to do cross-node parallel computation with Gaussian.

If you mean to disassemble one of the nodes, collect all the RAMs, install it on another node, and run the mission with more memory, it will speed up some type of calculation (if memory is the rate-limiting factor), but for missions like regular DFT with not too large molecule, it is usually not the case. And if you have multiple missions to run, using two nodes to run separate missions might be faster overall.

zako

First, thank you very much dear liyuanhe211 for replying.
I'm looking very hard for a method to accelerate my work, and i concluded that to reach my goal i need to use multiple nodes. I know it's impossible for me to buy Linda, but i read an article about the Parallelization using OpenMP, but i dont know how to benefit from it.
Please liyuanhe211 help me.

liyuanhe211

zako 发表于 2020-2-2 06:19
First, thank you very much dear liyuanhe211 for replying.
I'm looking very hard for a method to acc ...

For Gaussian program, as far as I know, the only cross-node parallelization model is via Linda. I'm afraid there is no other way to do it.

However, if you wish, you can try ORCA program. Not only it supports cross-node parallelization with openmpi (it's free. And I tested the cross-node parallel efficiency and get a pretty acceptable result), it also supports many acceleration methods well, like RI, DLPNO, which is very enabling for large molecules or for "poor" group, etc.

abin

OpenMP, for cpu cores to communicate and share data with each other, but the cpu cores must sit on the same mother board.
MPI, no mater, MPICH, openMPI, or intel MPI, or TCP-Linda as used by Gaussian, which is designed to communicate and share data among mother boards.

zako

Dear liyuanhe211 and Dear abin thank you very much for your helps and helpful informations.