Hi Zhaoxi,

Generally in HPC, hyper-threading is turned off. And you should use GPGPU card(s) in MD simulation.

It heavily depends, usually the overall benefit is about zero or negative and I saw it only positive for Intel CPUs (my experience of Intel HT vs AMD magny cours). For classical MD HT doesn't hurt too much, for QM which is usually memory bandwidth bounds, things are worse by far with HT.

With classical MPI parallelism, the number of communicators is also a huge factor as HT doubles the endpoints in that case. So for programs that allow a hybrid MPI/OpenMP parallelism, the ratio between both parallel layers is also a performance critical point and relies a lot on the hardware environment such as network interconnect.

As always in the HPC world, doing benchmarks is the best option. From my experience when people do benchmarks on a HT enabled system, they often neglect core binding which makes benchmark results very random since the threads can share more or less physical cores each time the test is started (when not using all cores). For example people do their tests using 1st: all available logical cores including the HT cores, or 2nd: half the available cores without binding. In this setup it is very likely that the 1st test produces the faster timings, since definetly all physical cores are involved, of course with some eventual penalty due the HT. the 2nd test on the other hand is most likely not using exclusively physical cores, but share them through logical HT cores.

So often even better results then the 1st test are obtained when doing the 2nd test but with binding processes/threads to physical cores only, so that no hyperthreads are used. It can sometimes be tricky to find the real layout of logical on the physical cores , f.i. /proc/cpuinfo can even be wrong due BIOS bugs. That's why it is safer usually to just switch off HT, since the benefit is only rarely considerable and only accessible by experts, while the potential of performance drops is so much higher for most users.

MD and QM calculations tend to rely on the floating point units, of which there are fewer in each physical core compared with arithmetic logic units. This leads to virtual threads trying to use the same unit so they have to keep waiting for their turn. As such it is quite common to switch off HT in HPCs that are used for floating point numerical work (MD, QM, CFD, FEA etc).