How to get best performance

From: SHIVAM TIWARI (t.shivam_at_iitg.ac.in)
Date: Tue Apr 26 2022 - 01:18:31 CDT

• Next message: Kevin Chan: "Compute atom/residue-based interactions on-the-fly"
• Previous message: Joey Farrell: "ABF Non-Uniform Zcount"
• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ] [ attachment ]

Dear users,
I am running a classical md simulation with around 1,20000 atoms with namd 2.14 on an hpc. I was trying to figure out the best configuration of nodes and cores for optimum performance by giving different combinations of cores and nodes and checking the benchmark results. when I gave 12 nodes * 10 tasks/node (120 cores) I got a performance of around 16 ns/day, and then I checked for other combinations also. I realized that 120 cores were giving me best performance. But when I again submitted my job with 120 nodes the performance reduced to 10 ns/day(why). So, I want to ask how can I get the best performance for my simulation? Also, I observed that after that after certain number of cores the performance saturates and we don't gain much by increasing the number of nodes and cores. Is there any way to increase the performance? or how do I know the best performance I can get for my system and how to get it?

thanks,
Shivam

• Next message: Kevin Chan: "Compute atom/residue-based interactions on-the-fly"
• Previous message: Joey Farrell: "ABF Non-Uniform Zcount"
• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ] [ attachment ]

This archive was generated by hypermail 2.1.6 : Tue Dec 13 2022 - 14:32:44 CST