When adding new VPP process there is a paradoxical performance behavior.

When doing performance testing consisting of sending a lot of API messages, VPP performance depends inversely on SLEEP_TIME_IN_SECONDS value (code below).

The performance is higher when SLEEP_TIME_IN_SECONDS low and lower when SLEEP_TIME_IN_SECONDS is high.

An example of this behavior is in file src/vnet/ip/rd_cp.c, rd_cp_process function: Performance was decreased after applying this patch: https://gerrit.fd.io/r/#/c/12521/, which decreases the sleep time from 10 microseconds to basically infinity.

static uword
example_process (vlib_main_t * vm, vlib_node_runtime_t * rt, vlib_frame_t * f)
{
  while (1)
    vlib_process_wait_for_event_or_clock (vm, SLEEP_TIME_IN_SECONDS);
  return 0;
}

VLIB_REGISTER_NODE (example_process_node) = {
    .function = example_process,
    .type = VLIB_NODE_TYPE_PROCESS,
    .name = "example-process",
};