Set our dedicated-per-numa value to 3 in the job to leave some room
for excluding cpu0 (which is special) in the next patch. Also try to
determine the appropriate number of dedicated cpus per test server
for cpu pinning by choosing half the minimum available amount on any
compute node.
Change-Id: If0bfd74f7ea9c3cfc7c2f21a939445ccb09e501d