On Mon, Oct 29, 2018 at 6:21 PM Ben Peart <peartben@xxxxxxxxx> wrote:
> @@ -114,6 +104,9 @@ static void preload_index(struct index_state *index,
> threads = index->cache_nr / THREAD_COST;
> if ((index->cache_nr > 1) && (threads < 2) &&
> git_env_bool("GIT_TEST_PRELOAD_INDEX", 0))
> threads = 2;
> + cpus = online_cpus();
> + if (threads > cpus)
> + threads = cpus;
> if (threads < 2)
> return;
> trace_performance_enter();
Capping the number of threads to online_cpus() does not always make
sense. In this case (or at least the original use case where we stat()
over nfs) we want to issue as many requests to nfs server as possible
to reduce latency and it has nothing to do with how many cores we
have. Using more threads than cores might make sense since threads are
blocked by nfs client, but we still want to send more to the server.
--
Duy
