Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

> c8g.8xlarge (32 core non-NUMA ARM64)

Requests are scheduled on half of these. Despite that, a plateau is hit after 8 threads? Is this a 16-core 32-thread type of a setup?

Also, consider redoing this in linear scale.

Edit: Oddly enough, no? 1 thread per core as per https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/cpu-opti...



> a plateau is hit after 8 threads?

Most of the graphs plateau around 6 threads, for pretty much all the caches under test. I wonder if there is some interesting architectural issue with cache-sharing on this particular platform?


I guess memory controller behavior; especially if it's not set up for parallelism IOPS but for single-threaded throughout (channel interleaving).


Thanks for the feedback. The plateauing is due to a taskset misconfiguration. I fixed it and reran the benches. They look much better now.


g = graviton which doesn't support SMT so 1 vcpu is 1 full core




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: