I'm trying to patch the 2.6.18 kernel for our company needs, which requires an *extremely* low latency for handling incoming UDP packets. One of the main issues I'm trying to address is scheduling latency, which AFAIK can be dramatically reduced by adding preemption points.
I'm using the preemption compilation flag, but would like to check the possibility of manually adding more points in the code. The special case here is that the system is under a relatively light load so throughput is not important, yet every microsecond of latency while handling incoming packets counts.
Does anyone happen to have tips relevant for me? I'm mainly interested in ideas regarding the "correct" places to add more preemption calls, yet any interesting insights will be helpful :)