Addressing Performance Dip in Long-Term NetLogo Simulations

[Tim Verlaan]

Hi All,

I'm using a slurm HPC infrastructure to run NetLogo models on and am experiencing a performance issue that occurs consistently after approximately 4'400'000 ticks.

The model starts to exhibit a noticeable and, as observed, empirically predictable slowdown. Despite having increased RAM from 30 GB to 60 GB per model and despite efforts to manage NetLogo resources, by clearing the shortest path cache every 100'000 ticks and handling data exports in batches, clearing lists at the same frequency, the issues remains (and still occur after 4'400'000 ticks). I cannot identify any elements of my code nor of the input data (I have virtual police units dispatch and handle real historical calls for service) that would explain this. Also, I doubt it has to do with any thing HPC side as it is so consistent on the 4'400'000 regardless of when either of the models reach that value of ticks.

Here are my questions:

1. Has anyone faced similar performance issues in long-term simulations in Netlogo? Could any known issues be responsible for a dip at a specific ticks points?
   - If so, how did you address them?
2. Are there any recommended strategies or adjustments to combat such performance dips in extensive simulations?

Any advice or insights would be greatly appreciated!

Below, you'll find a figure where a black line indicates the minimal progression needed for the models to all complete a scope of simulated evolution within the 5-day time frame quota, allocated by the HPC.

P.S.: I have ways to bring random (green) to the performance of stationary. However, as you can see, eventually all will be too slow...

P.P.S.: a level of transparency of the graphed lines is governed by the number of agents traversing the street network in the models (so the more transparent, the less agents, the better the performance - although they all suffer the hit at 4'500'000 ticks )

P.P.P.S.: Also see my StackOverflow post here: https://stackoverflow.com/questions/78601389/addressing-performance-dip-in-long-term-netlogo-simulations

[Aaron Andre Brandes]

Hi Tim,

Running NetLogo on my MacBook Pro I have found that large models run for many ticks can experience slowdowns due to memory usage and garbage collection at times, but not at predictable tick numbers. I also haven’t run them for as long as you have.

 

If you haven’t read it you should check out

 

Improving Execution Speed of Models Implemented in NetLogo

Steven Railsback1, Daniel Ayllón2, Uta Berger3, Volker Grimm2, Steven Lytinen4, Colin Sheppard5, Jan Thiele6

 

Journal of Artificial Societies and Social Simulation 20(1) 3, 2017

Doi: 10.18564/jasss.3282 Url: http://jasss.soc.surrey.ac.uk/20/1/3.html

 

You may be able to improve performance by making changes to your model.

 

Please include bu...@ccl.northwestern.edu bu...@ccl.northwestern.edu on any followup emails to assure the attention of our NetLogo developers.

- Aaron

-- 

Aaron Brandes, Software Developer

Center for Connected Learning and Computer-Based Modeling

 

 

From: netlog...@googlegroups.com <netlog...@googlegroups.com> on behalf of Tim Verlaan <t.v.v...@gmail.com>
Date: Tuesday, June 11, 2024 at 9:00 AM
To: netlogo-users <netlog...@googlegroups.com>
Subject: [netlogo-users] Addressing Performance Dip in Long-Term NetLogo Simulations

Hi All,

I'm using a slurm HPC infrastructure to run NetLogo models on and am experiencing a performance issue that occurs consistently after approximately 4'400'000 ticks.

The model starts to exhibit a noticeable and, as observed, empirically predictable slowdown. Despite having increased RAM from 30 GB to 60 GB per model and despite efforts to manage NetLogo resources, by clearing the shortest path cache every 100'000 ticks and handling data exports in batches, clearing lists at the same frequency, the issues remains (and still occur after 4'400'000 ticks). I cannot identify any elements of my code nor of the input data (I have virtual police units dispatch and handle real historical calls for service) that would explain this. Also, I doubt it has to do with any thing HPC side as it is so consistent on the 4'400'000 regardless of when either of the models reach that value of ticks.

Here are my questions:

1.      Has anyone faced similar performance issues in long-term simulations in Netlogo? Could any known issues be responsible for a dip at a specific ticks points?

- If so, how did you address them?

2.      Are there any recommended strategies or adjustments to combat such performance dips in extensive simulations?

Any advice or insights would be greatly appreciated!

Below, you'll find a figure where a black line indicates the minimal progression needed for the models to all complete a scope of simulated evolution within the 5-day time frame quota, allocated by the HPC.

P.S.: I have ways to bring random (green) to the performance of stationary. However, as you can see, eventually all will be too slow...

P.P.S.: a level of transparency of the graphed lines is governed by the number of agents traversing the street network in the models (so the more transparent, the less agents, the better the performance - although they all suffer the hit at 4'500'000 ticks )

P.P.P.S.: Also see my StackOverflow post here: https://stackoverflow.com/questions/78601389/addressing-performance-dip-in-long-term-netlogo-simulations

 

--
You received this message because you are subscribed to the Google Groups "netlogo-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to netlogo-user...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/netlogo-users/8c5949fa-a419-47bf-b597-347aefe8c376n%40googlegroups.com.