HI,
I have just started utilizing boost fibers and am completely confused about its conditional variable concurrency behavior.
Boost fibers provide the following example for using conditional variables:
void wait_for_data_to_process() {
{
std::unique_lock< boost::fibers::mutex > lk( mtx);
while ( ! data_ready) {
cond.wait( lk);
}
} // release lk
process_data();
}
void prepare_data_for_processing() {
retrieve_data();
prepare_data();
{
std::unique_lock< boost::fibers::mutex > lk( mtx);
data_ready = true;
}
cond.notify_one();
}
But when we do the above we see that the fibers end up running into std::unique_lock waits in system calls sched_yield and futex. We are possibly invoking above concurrency primitives among multiple OS threads. We want to avoid system calls in the fibers.
strace outputs are as below:
futex(0x7ff4f4005160, FUTEX_WAKE_PRIVATE, 1) = 0
> /usr/lib64/libc.so.6(__GI___lll_lock_wake+0x17) [0x87777]
> /usr/lib64/libc.so.6(__GI___pthread_mutex_unlock_usercnt+0x94) [0x8f2d4]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::algo::round_robin::notify()+0x24) [0x73f4]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::context::wake(unsigned long)+0x29) [0xaac9]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::wait_queue::notify_one()+0x40) [0xbca0]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::mutex::unlock()+0x31) [0xc151]
sched_yield() = 0
> /usr/lib64/libc.so.6(__sched_yield+0xb) [0xf3d5b]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::detail::spinlock_ttas::lock()+0xd4) [0x9044]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::mutex::lock()+0x28) [0xc198]
Apparently even scheduling a fiber seems to run into lock waits, not sure why scheduling a fiber would block:
sched_yield() = 0
> /usr/lib64/libc.so.6(__sched_yield+0xb) [0xf3d5b]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::detail::spinlock_ttas::lock()+0xd4) [0x9044]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::scheduler::remote_ready2ready_()+0x33) [0xd353]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::scheduler::dispatch()+0x50) [0xd4e0]
> /usr/local/lib/libboost_fiber.so.1.85.0(boost::fibers::dispatcher_context::run_(boost::context::fiber&&)+0xd) [0xb46d]
> /usr/local/lib/libboost_fiber.so.1.85.0(void boost::context::detail::fiber_entry<boost::context::detail::fiber_record<boost::context::fiber, boost::fibers::stack_allocator_wrapper, std::_Bind<boost::context::fiber (boost::fibers::dispatcher_context::*(boost::fibers::dispatcher_context*, std::_Placeholder<1>))(boost::context::fiber&&)> > >(boost::context::detail::transfer_t)+0x3a) [0xb4ca]
> /usr/local/lib/libboost_context.so.1.85.0(make_fcontext+0x2e) [0x117e]
> No DWARF information found
We want to avoid all system calls in the fibers to avoid their performance impact, we are confused why the examples provided by boost fibers are running into system calls.
Any help is greatly appreciated ...
regards
-Amit
