[dist1ll]

RPS meaning reads or writes? What's the distribution of message sizes, and how large is your total dataset? What specs (core count, NIC) did each node have?

I'm asking because without this info, RPS is not a particularly useful metric. As an extreme example, if your dataset is <1MB, you could likely serve read-heavy requests from your SmartNIC's dcache at close to line rate.

[sethammons]

It's been a number of years since I worked on it. I can try to answer your questions. The calls were nearly entirely INCR calls against keys that were typically around 150 bytes long and app logic was based on the return values. I believe each node took up around 20GB of memory when saturated. Redis uses a single CPU, so core count shouldn't matter, no? I'm not entirely sure what our NIC situation was, but it was tuned to handle gigabit traffic.

[brazzledazzle]

Did you tune the file descriptor limits?

[sethammons]

yeah, good call out. We did. One of the things a lot of folks stub their toe on when starting to scale up