Simulink simulation slows down drastically after first run following reboot

Hi! I am having an issue with a simulink simulation. The first time I run it after a windows reboot, the simulation rate is somewhere around 0.7 simulation seconds / real seconds. When I simulate again, the simulation rate drops to about 0.1 simulation seconds / real seconds. Subsequent runs slow down further and stabilize somewhere around estimated 0.02 simulation seconds / real seconds. For a simulation that needs to simulate 150 seconds to be meaningful, this is far too slow. Since I cannot publicly post this simulation, I will describe properties of the simulation, as well as observations I have made regarding this issue.

Description of simulation:

• Heavily using Simscape Power Systems -> Specialized Technology
• Discrete fixed-step solver (auto-select), no continuous states
• Rapid-accelerator mode (though this also seems to happen when using normal mode)
• most parameters are defined using a struct "par" in the base workspace. Simulink blocks are parameterized by accessing this struct (e.g. a constant value would be set by entering "par.someConstantValue").
• There is a use of masked and variant subsystems

Computer info:

• Windows 10 64-bit
• 32GB of ram (absolutely not limiting for the case I'm describing here)
• Core i7-8700K
• Geforce 1060

Observations:

• This issue likely exists/has become severe since I ran this simulation overnight for a long time with lots of output data. It probably was too much data. The computer had crashed the next day I came back to it. This was using Matlab 2017b
• I thought I might as well update to Matlab 2018a afterwards. No difference for this issue, though.
• Another simulation shows the same slow-downs. It is based on the Simscape Power Systems example "Thyristor-Based HVDC Transmission System (Detailed Model)". As soon as I change the parameterization to the struct-model described above, simulation performance drops by a few orders, but remains stable. Therefore, I am not really sure if this is the core of the problem.

I do not know where to look and what else to try. I have sunk a lot of time into this already and need to become productive again. I am going to test this on another computer, as soon as I can. Does anyone have any help/clues/support?

Thank you very much! MP

Edit: please see second answer posted for further information.

Answer 1

[MP]

I did some systematic testing with lots of windows/matlab/simulation restarting tonight (18 test runs). Observations are as follows:

1. The slprj and mexw64 do not seem to be connected to this issue
2. I get one fast (0.7) run after a windows reboot. The next run will be slower (0.1, usually stabilizes somewhere around there for consecutive runs without any restarts)
3. If I alter my parameter struct (par), the simulation speed will drop to super-slow (the 0.02 mentioned above). This can ONLY be reset by a windows restart. Even restarting matlab does not change this
4. I saw that a new nvidia graphics driver was released 3 days ago (version 391.35). So I updated from the previous version (391.24). After that, the simulation slowing was a little less (maybe settling between 0.03-0.05 instead of 0.02). However, everything else is still unchanged, especially the way it slows down.

Answer 2

[MP]

I cannot reproduce this problem on a different computer (unfortunately much slower), hence it must be related to some OS, driver or configuration thing on this computer.

Nevertheless, I switched to the 'Accelerator' mode and don't seem to be having this issue. I will continue using this mode from now on.

Answer 3

[Ilham Hardy]

Since you run the model in Rapid Accelerator mode, have you tried deleting the slprj and yourmodelname.mexw64?

[MP]

Hi Illham, thank you for your reply. I did try this and it has no effect on the situation. I will post a more elaborate answer below in a few moments.