I don't know anything about the Linux scheduler, but if one core out of 96 is busy chasing after those ghost allocations, why would a network thread not use one of the other 95 cores, and instead be stuck for seconds? Unless that one core is blocking the entire kernel, but then that would be a lot more visible instantly.
I don't know anything about the Linux scheduler, but if one core out of 96 is busy chasing after those ghost allocations, why would a network thread not use one of the other 95 cores, and instead be stuck for seconds? Unless that one core is blocking the entire kernel, but then that would be a lot more visible instantly.