Before we've jitted the first bit of the TC, having so many
threads leads to a pathological situation: we accept a bunch of
requests (135) and have all of them running in the interpeter,
starving the JIT thread that is trying to make them all go faster.
(Perf record of a machine in this situation shows the vast majority of
the time in dispatchImpl, SrcDB::find, and enterTC). This diff tries
to solve this by limiting the number of threads to the number of cores
for the first N requests.
Jordan DeLong
230d2ae5e6