On 2021/03/24 01:09, Dario Faggioli wrote:
This has been very useful to me, when looking at large boxes, with more than one hundred CPUs, as looking at /proc/interrupts is a nightmare there. ... These seems to be focusing on understanding the "IRQ layout" of the system, and working with affinities, which are also really interesting and useful things indeed.
Bingo -- I had no clue what was coming in where, and needed something to summarize what I was seeing -- that and was trying to optimize interrupt servicing for 10-20Gb ethernet. Conversely, the most cpu's on a box I have is 12 -- so less need for something that can aggregate 100. The box w/12 has 6/NumaNode (actual cores, not threads), so at least tying devs to 1 node or the other can give small benefits, but best improvement on my network are 9k packets (v 1.5k) and using internal-lines that don't need encryption, yielding 350-650MB/s on reads and about 360MB/s on writes. Programming things like that in shell can be nice when it works, as those are ~9 years old since last touched. Cheers! -l