gcc version

While we won't change the default toolchain in Amazon Linux 2 (for all the compatibility reasons you can imagine), additional versions that could sit side-by-side are certainly on the list of things we could possibly do. This kind of task naturally sits in a list of competing priorities for our attention, so I can't promise anything.

Are there any specific packages (apart from the kernel) that you're looking at a newer toolchain for, and that a newer toolchain would provide a large boost?

We have backported some changes to GCC in Amazon Linux 2 in order to help with performance on Graviton 2 based instances, so our 7.3.1 isn't really the equivalent of upstream GCC 7.3.1 from a performance point of view.

Hi @Stewart,
I have installed a side-by-side version of the GNU compilers in several of my AMIs. As a matter of fact, your AWS-parallelcluster team does so in their releases. However, some apps just won't compile if they detect a discrepancy with the compiler used to build the kernel. An example is the CUDA toolkit, which is a bit on the touchy side. My current problem is not with CUDA but with WRF (maybe you are not familiar with this app) but it also complains about this disagreement. WRF developers and users have detected performance differences based on compilers and versions. However, I cannot point out which modules are responsible for performance gain so I'm unsure if the differences between the upstream GCC version and the ALinux version would make any difference.

Edited by: afernandezody on Jun 29, 2020 11:23 AM

CUDA is a bit special as it will end up building shim kernel modules in order to load in the proprietary NVIDIA module.

I'm not personally familiar with WRF. I would be interested to know what it is and if there's any good benchmarks we could look at.

WRF stands for Weather Research and Forecasting Model (you can find more info at https://www.mmm.ucar.edu/weather-research-and-forecasting-model). Notice that it's meant to be run on clusters so you'll need, at a minimum, a large instance.

Thanks, looks interesting. Do you happen to know if there's a good data set and operation to run that would help us evaluate performance?

Hi Stewart,
Sorry that I didn't notice your latest post earlier. I'm using some of the benchmarks described at https://www2.mmm.ucar.edu/wrf/WG2/bench/. The app is complex and there are countless different scenarios, but the said website should be a good starting point. You can also gather more info and contact the developers via the official forum (https://forum.mmm.ucar.edu/phpBB3/viewforum.php?f=25&sid=1e9152b410b1369921fb9b0ebf65246c). Hope this helps.
AF