Tech support for PSSP on IBM SP2/ aix ...
Customer called:and complained about throughput being 16% too low over the css0 interface. Disassembly of the locking code in the kernel showed that the CPU was doing a cache flush before transmission of a packet which it was not supposed to for this CPU type. This was traced back to a bitmask set in a struct in if.h which was slightly wrong.
Ok, calculate the correct bitmask, go to the kernel debugger, find the network interface address for css0, count 92 bytes, and patch the right value into the kernel.
Rt voila, 15% more throughput, customer was happy. Took us a while to find this bug though, and the how-to was accompanied by a disclaimer. Poughkeepsie, our level3 support, provided an official fix.