Thanks for having me!

Hi everyone!

Hello everyone, thank you for joining me this month at HBC Cafe.

My name is Anna, and this afternoon we will talk about the molecular dynamics package Gromics.

We will start with a short overview of the history,

then I will share some of our benchmarking results.

And in the end, we will see how to use Gramics on GPUs and CPUs.

Gramics was first published in 1995.

And its functionality is based on another molecular dynamics

simulation package namely the Gromos.

a lot of the functions were copied but also a lot of new algorithms were added.

The word GROMACS is an acronym for Groningen Machine for Chemical Simulation and this was actually

a computer, a physical computer, a processor ring that ran the MD simulations.

In the beginning there were only two force fields available, but the designers and developers were

able to run, well, one nanosecond in three days. But for the beginning, this was totally cool.

Then we go on, new versions were published and another force field was added. And what's actually

quite nice about the new developments is that the developers used the multimedia instructions on the

processors that are there for lightning processing in games and could use this feature to calculate

the distances, the inter-particle vectors between pairwise distances from particle coordinates.

I'm not going into detail what the developers then added. They added OpenMP multi-threading

and ported the code to GPUs and even more developments on GPUs transitioned to a different

language and in the end we end up with version 2022 that improves the direct communication on the GPUs.

That's all about the history. We can offload a lot of calculations from the MD steps

to the GPU and by that the performance can be increased. Now let's look at some of the benchmarks.

The first case I'm going to show you is a standard all-MD simulation. It's just a protein in a membrane

and with explicit water. All of the possible calculations are offloaded to the GPU.

These are the numbers that I produced and we can see that by quadrupling the number of the GPUs

in the first line for GROMiX 2019 we get up to 96 more nanoseconds per day. This is about an

performance increase of 72% than with only one GPU. But if we look at the line with GROMiX 2021.1

we get for one GPU 233 nanoseconds per day which is actually the same as with GROMiX 2019 on four GPUs.

So the take-home message here is use the newer version if possible because we can get the same

increase in performance when we use the newest version as compared to just using multiple GPUs.

Now the reason why the performance on two GPUs is so much better than on one GPU is because

the update functionality only works with updates groups and with hydrogen

bond constraints and not all bond constraints. So this is something I'm going to talk about later

but it is something to consider to only use constraints on two GPUs. So,

I'm going to talk about later but it is something to consider to only use constraints on hydrogen

bonds and not on all other bonds. The newest version is not completed yet so the

performance data is not that good at the moment so I would recommend to use the

GROMIX version 2021.5 at the moment but we need individual tests to gain an impression

of the actual performance per system. This is the same case, it's also the offload of every

calculation on GPU and what we can see here is that the newest generation of GPUs

gives the best overall performance because the GPUs also improved the architecture,

there are more cores available, the communication is better. What we can see here again is that when

we increase the number of GPUs we might only gain two nanoseconds of performance increase

and this doesn't make sense, right? Why should we use more GPUs just to get two nanoseconds?

Below the nanosecond is the parallel efficiency and this is the ratio of the speed up which

quantifies how much faster we can compute with more than one device and it tells us

which fraction of the resources is actually used for computation. So in this case a fourth of the

hardware is used. We use four GPUs and we get the same performance as for one GPU. So even if we use