Suggestion of best AWS Instances for CP-SAT

[Nara Torres Moreira]

Hello, I'm working with CP-SAT models to solve large scheduling problems in the industry.

Currently my team has a few applications deployed, which so far share a m5.4xlarge AWS instance. However, I noticed that the performance with this instance is not good enough for a new model that has to be deployed. The tests were done using 8 CPUs for the model, and the performance was much worse than in my local machine also using 8 CPUs (MacBook Pro, specs: https://support.apple.com/en-gb/111902). In terms of memory, the peak was 12 GB being used by the solver (short peak).

Now we are talking about changing the AWS instance, and I was wondering if anyone here can recommend which instances are more adequate/compatible for running large CP-SAT models. In general the more CPUs the better (although there might be a limit for the continuous improvement?), but I know there are other specs that can influence the performance.

Thank you very much,

Nara

[Laurent Perron]

I will not comment on AWS' offering. But server CPUs are around 2-3x slower than apple chips.

And this does not include hyperthreading or equivalent that makes half the core way less performant (note that this is also true for energy efficient apple cores).

Note that I answered a related question here: https://github.com/google/or-tools/discussions/4751

Laurent Perron | Operations Research | lpe...@google.com | (33) 1 42 68 53 00

--
You received this message because you are subscribed to the Google Groups "or-tools-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to or-tools-discu...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/or-tools-discuss/f5b2dee9-7770-4596-85e1-a23e0e04c30bn%40googlegroups.com.

[Nara Torres Moreira]

Thank you for the quick response and for the info!

[grego...@gmail.com]

I think it is more about suitable hardware for cp-sat solver. Generally it is good to have strong single core and multi-core performance, something like threadripper or epyc f variant with higher frequencies compared to standard servers.

Good start is to probably measure how number of search workers influence performance of your application. Sometimes less workers is similar or better due to higher frequencies, sometimes more workers is more important. 

m5.4xlarge AWS instances use very old cpu so it is no surprise that m1 pro is better. Based on my experience with m2 pro I can say that multi-core performance is not very good probably due to thermal limits. cp-sat solver is rare example when you can hear macbook pro fan :-)

Jan