MAPDL launching time analysis

[germa89]

Summary

Analyse the time spent by MAPDL to launch in different configurations.

Goal

For the CICD purposes (mainly pool testing) I would like to reduce the time to launch MAPDL to the minimum. For that reason, I am investigating different configurations for this.

Setup

MacOS M2 2022 Ventura 13.5
Docker desktop 4.21.1 (114176)
Docker Engine: 24.0.2
Docker image based on Ubuntu 22.04.2 LTS and MAPDL v22.2

Configurations

We are checking the impact of SMP/DMP, number of processor and amount of RAM in the launching time of MAPDL.

• -m specifies the amount of memory which MAPDL will start with. In MB by default
• -np specify the amount of processors.
• -smp is the flag that activates the Shared Memory Processing
• -dmp is the flag tat activated Distributed Memory Processing. Uses MPI, it is more time consuming.

Results

Using smp -np 1 -m 100 can reduce the launching time 50% which is good enough.

Detailed console tests

Console output without grpc exiting cleanly (using /EXIT)

root@8db57f17c0b5:~/jobs# time $(ansys <<< "/exit" > log.log)

real    0m22.149s
user    0m13.836s
sys     0m3.557s
root@8db57f17c0b5:~/jobs# time $(ansys -smp  <<< "/exit" > /dev/null )

real    0m6.955s
user    0m2.499s
sys     0m0.450s
root@8db57f17c0b5:~/jobs# time $(ansys -smp -np 2 <<< "/exit" > /dev/null )

real    0m10.250s
user    0m2.635s
sys     0m0.426s
root@8db57f17c0b5:~/jobs# time $(ansys -smp -np 1 <<< "/exit" > /dev/null )

real    0m7.335s
user    0m2.527s
sys     0m0.423s
root@8db57f17c0b5:~/jobs# time $(ansys -smp -np 1 -m 100 <<< "/exit" > /dev/null )

real    0m7.154s
user    0m2.699s
sys     0m0.430s
root@8db57f17c0b5:~/jobs# time $(ansys -smp -np 1 -m 50 <<< "/exit" > /dev/null )

real    0m7.060s
user    0m2.592s
sys     0m0.429s
root@8db57f17c0b5:~/jobs# time $(ansys -smp -np 1 -m 20 <<< "/exit" > /dev/null )

real    0m7.450s
user    0m2.567s
sys     0m0.439s

Console output with grpc

The logic to time this is a bit more complex:

time $(ansys -grpc <<< echo -ne '\n' | tee /dev/tty | (grep -q "Server listening" && pkill -9 ansys.e))

I had to pipe Ansys output to /dev/tty so I can read it after with grep -q (-q to exiting the process when a match is found), then after I found out a match, I kill the process.

root@8db57f17c0b5:~/jobs# time $(ansys -grpc <<< echo -ne '\n' | tee /dev/tty | (grep -q "Server listening" && pkill -9 ansys.e))
real    0m13.765s
user    0m1.946s
sys     0m0.417s

root@8db57f17c0b5:~/jobs# time $(ansys -grpc -smp  <<< echo -ne '\n' | tee /dev/tty | (grep -q "Server listening" && pkill -9 ansys.e))  
real    0m10.686s
user    0m2.622s
sys     0m0.451s

root@8db57f17c0b5:~/jobs# time $(ansys -grpc -smp -np 2 <<< echo -ne '\n' | tee /dev/tty | (grep -q "Server listening" && pkill -9 ansys.e))  
real    0m9.656s
user    0m2.693s
sys     0m0.484s

root@8db57f17c0b5:~/jobs# time $(ansys -grpc -smp -np 1 <<< echo -ne '\n' | tee /dev/tty | (grep -q "Server listening" && pkill -9 ansys.e))  
real    0m9.151s
user    0m2.647s
sys     0m0.434s

root@8db57f17c0b5:~/jobs# time $(ansys -grpc -smp -np 1 -m 100 <<< echo -ne '\n' | tee /dev/tty | (grep -q "Server listening" && pkill -9 ansys.e)) 
real    0m8.206s
user    0m2.657s
sys     0m0.467s

root@8db57f17c0b5:~/jobs# time $(ansys -grpc -smp -np 1 -m 50 <<< echo -ne '\n' | tee /dev/tty | (grep -q "Server listening" && pkill -9 ansys.e))  
real    0m8.213s
user    0m2.695s
sys     0m0.434s

root@8db57f17c0b5:~/jobs# time $(ansys -grpc -smp -np 1 -m 20 <<< echo -ne '\n' | tee /dev/tty | (grep -q "Server listening" && pkill -9 ansys.e))  
real    0m8.488s
user    0m2.697s
sys     0m0.437s

[FredAns]

Hi @germa89,

We can understand why launching MAPDL with minimal amount of resources can speed-up the starting step. And perhaps we could expose some few configurations, wrt what the user intend to do.

Using SMP is ok to solve small size problems, or just to do some R&D.
But when we deal with medium to big size models, the time you saved to start MAPDL will be nothing compared to the additional time you will need to solve.
I would say for models > 100.000 equations, it really worths to switch to DMP.

Default configuration in Mechanical is to start MAPDL in DMP with 4 cores.
If the machine we are running on have more cores, it may be good to increase this number, (but not all cores).

We could have some predefined configs depending what the user needs, Is it what you have in mind ?

[germa89]

I have nothing in mind particularly. But we could suggest number of processors/memory depending on the machine. It should not be difficult to implement.

Mostly I wanted this "discussion" to gather some opinions, advices, tricks with regards to launching MAPDL. Indeed your comment is very valuable.

Thank you!