Improve EventCount used by the non-blocking threadpool.

The current algorithm requires threads to commit/cancel waiting in order
they called Prewait. Spinning caused by that serialization can consume
lots of CPU time on some workloads. Restructure the algorithm to not
require that serialization and remove spin waits from Commit/CancelWait.
Note: this reduces max number of threads from 2^16 to 2^14 to leave
more space for ABA counter (which is now 22 bits).
Implementation details are explained in comments.
This commit is contained in:
Rasmus Munk Larsen
2019-02-22 13:56:26 -08:00
parent 0b25a5c431
commit 6560692c67
3 changed files with 110 additions and 93 deletions

View File

@@ -374,11 +374,11 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
eigen_plain_assert(!t->f);
// We already did best-effort emptiness check in Steal, so prepare for
// blocking.
ec_.Prewait(waiter);
if (!ec_.Prewait()) return true;
// Now do a reliable emptiness check.
int victim = NonEmptyQueueIndex();
if (victim != -1) {
ec_.CancelWait(waiter);
ec_.CancelWait();
if (cancelled_) {
return false;
} else {
@@ -392,7 +392,7 @@ class ThreadPoolTempl : public Eigen::ThreadPoolInterface {
blocked_++;
// TODO is blocked_ required to be unsigned?
if (done_ && blocked_ == static_cast<unsigned>(num_threads_)) {
ec_.CancelWait(waiter);
ec_.CancelWait();
// Almost done, but need to re-check queues.
// Consider that all queues are empty and all worker threads are preempted
// right after incrementing blocked_ above. Now a free-standing thread