static void do_parallel_map(Stopwatch& sw) {
// Create a bunch of subcontexts from here and put them in a map
auto all = create();
::ContextMap<size_t> mp;
for (size_t i = N; i < all.size(); i++)
mp.Add(i, all[i]);
// Now the threads that will make progress hard:
auto proceed = std::make_shared<bool>(true);
for (size_t i = N; i--;)
std::thread([proceed, mp] {
while (*proceed)
for (const auto& cur : mp)
;
}).detach();
auto cleanup = MakeAtExit([&] { *proceed = false; });
std::this_thread::sleep_for(std::chrono::milliseconds(100));
// Enumerate the map while iteration is underway
sw.Start();
for (auto& cur : mp)
;
sw.Stop(n);
}
TEST_F(CoreContextTest, AppropriateShutdownInterleave) {
// Need both an outer and an inner context
AutoCurrentContext ctxtOuter;
AutoCreateContext ctxtInner;
// Need to inject types at both scopes
AutoRequired<ExplicitlyHoldsOutstandingCount> outer(ctxtOuter);
AutoRequired<ExplicitlyHoldsOutstandingCount> inner(ctxtInner);
// Start both contexts up
ctxtOuter->Initiate();
ctxtInner->Initiate();
// Now shut down the outer context. Hand off to an async, we want this to block.
std::thread holder{
[ctxtOuter] {
ctxtOuter->SignalShutdown(true);
}
};
auto holderClean = MakeAtExit([&holder] { holder.join(); });
// Need to ensure that both outstanding counters are reset at some point:
{
auto cleanup = MakeAtExit([&] {
outer->Proceed();
inner->Proceed();
});
// Outer entry should have called "stop":
auto future = outer->calledStop.get_future();
ASSERT_EQ(
std::future_status::ready,
future.wait_for(std::chrono::seconds(5))
) << "Outer scope's OnStop method was incorrectly blocked by a child context member taking a long time to shut down";
}
// Both contexts should be stopped now:
ASSERT_TRUE(ctxtOuter->Wait(std::chrono::seconds(5))) << "Outer context did not tear down in a timely fashion";
ASSERT_TRUE(ctxtOuter->IsQuiescent()) << "Quiescence not achieved by outer context after shutdown";
ASSERT_TRUE(ctxtInner->Wait(std::chrono::seconds(5))) << "Inner context did not tear down in a timely fashion";
}
Benchmark PriorityBoost::CanBoostPriority(void) {
AutoCurrentContext ctxt;
// Create two spinners and kick them off at the same time:
AutoRequired<JustIncrementsANumber<ThreadPriority::BelowNormal>> lower;
AutoRequired<JustIncrementsANumber<ThreadPriority::Normal>> higher;
ctxt->Initiate();
#ifdef _MSC_VER
// We want all of our threads to run on ONE cpu for awhile, and then we want to put it back at exit
DWORD_PTR originalAffinity, systemAffinity;
GetProcessAffinityMask(GetCurrentProcess(), &originalAffinity, &systemAffinity);
SetProcessAffinityMask(GetCurrentProcess(), 1);
auto onreturn = MakeAtExit([originalAffinity] {
SetProcessAffinityMask(GetCurrentProcess(), originalAffinity);
});
#else
// TODO: Implement on Unix so that this benchmark is trustworthy
#endif
// Poke the conditional variable a lot:
AutoRequired<std::mutex> contended;
for(size_t i = 100; i--;) {
// We sleep while holding contention lock to force waiting threads into the sleep queue. The reason we have to do
// this is due to the way that mutex is implemented under the hood. The STL mutex uses a high-frequency variable
// and attempts to perform a CAS (check-and-set) on this variable. If it succeeds, the lock is obtained; if it
// fails, it will put the thread into a non-ready state by calling WaitForSingleObject on Windows or one of the
// mutex_lock methods on Unix.
//
// When a thread can't be run, it's moved from the OS's ready queue to the sleep queue. The scheduler knows that
// the thread can be moved back to the ready queue if a particular object is signalled, but in the case of a lock,
// only one of the threads waiting on the object can actually be moved to the ready queue. It's at THIS POINT that
// the operating system consults the thread priority--if only thread can be moved over, then the highest priority
// thread will wind up in the ready queue every time.
//
// Thread priority does _not_ necessarily influence the amount of time the scheduler allocates allocated to a ready
// thread with respect to other threads of the same process. This is why we hold the lock for a full millisecond,
// in order to force the thread over to the sleep queue and ensure that the priority resolution mechanism is
// directly tested.
std::lock_guard<std::mutex> lk(*contended);
std::this_thread::sleep_for(std::chrono::milliseconds(1));
}
// Need to terminate before we try running a comparison.
ctxt->SignalTerminate();
return Benchmark {
{"Low priority CPU time", std::chrono::nanoseconds{lower->val}},
{"High priority CPU time", std::chrono::nanoseconds{higher->val}},
};
}
static void do_parallel_enum(Stopwatch& sw) {
AutoCurrentContext ctxt;
auto all = create();
// Create threads which will cause contention:
auto proceed = std::make_shared<bool>(true);
for (size_t nParallel = N; nParallel--;) {
std::thread([proceed, all, ctxt] {
while (*proceed)
for (auto cur : ContextEnumerator(ctxt))
;
}).detach();
}
auto cleanup = MakeAtExit([&] { *proceed = false; });
std::this_thread::sleep_for(std::chrono::milliseconds(100));
// Perform parallel enumeration
sw.Start();
for (auto cur : ContextEnumerator(ctxt))
;
sw.Stop(n);
}
