Description
When testing various options, I noticed a problem where a kernel spin-lock would consume a noticeable amount of CPU (like, half of all CPUs). This was on a dual-socket 6-core system (so total of 24 logical CPUs, 12 physical and 12 HT). It would also peg one SIRQ handled (for /0) at close to 100%.
I tracked this down to disabling accept threads. I don't know why this is happening, but it clearly shows that keeping accept-threads enabled is better. The downside is that accept-threads currently experiences higher latencies.
Attachments
Issue Links
- relates to
-
TS-3442 Delay seen in handling the client's request while running with accept thread
- Open