Efficiency of i-pi

[Kris Mu]

Dear all,

After using i-PI for a while, I have a certain impression. When my driver performs calculations using a GPU, such as with a machine learning force field, the speed of MD actually depends on the performance of both the CPU and the GPU. For example, take two machines: one has a very good CPU but an average GPU, while the other is the opposite. In the end, the running speed of i-PI on both machines might be similar, and for smaller systems, it might even be entirely dependent on the CPU speed. I think this could be because the server process of i-PI is a Python program running on the CPU. Is my impression reasonable?

Many thanks for clearing my confusion！

[Michele Ceriotti]

We discuss in detail the overhead of running i-PI in the publication that accompanies the latest release,

https://doi.org/10.1063/5.0215869

The bottom line is that yes, the overhead can be substantial but only when you either run a very complicated

setup on the MD side, or use a very cheap potential. For simple MD and a medium-size system of ~1000 atoms,

you should expect an overhead below 1ms/step. If you see overheads much larger than this, then:
1. make sure you use a version > 3.0, we sped up everything by a lot for the last release

2. reduce I/O, printing positions or velocities at each step is usually unnecessary, and takes quite some time

3. make sure that the socket communication is not unnecessarily slow, e.g. use a unix-domain socket

if you run on a single node.

Hope this helps

Michele