Why is the rtt measured here on canon lake so asymmetric? I would have expected symmetry along the leading diagonal.

And why is core 0 magical on that machine?

IIRC, core 0 is the one that tends to service processor interrupts, which may mean that the data is unusually noisy in a way that tends to underestimate the RTT latency time.

There’s a similar weird artifact at index 64.

Core 64 is the hyper-threaded pair of core 0.

Not on that machine. That would only be true on a 128-thread system, but this one has 96 threads.

Speculation, but could it be because of a ring topology?

This might be a nitpick, but it seems like already the intro sentence isn't factually accurate:

> Over the past 10-15 years, per-core throughput slowed down

I assume that the author meant that the acceleration of per-core throughput over time has slowed down, not the throughput itself.

> Over the past 10-15 years, per-core throughput slowed down

throughput of what?