high cpu usage in boost::asio::io_service::run - c++

Question

I have encountered a strange problem with boost::asio::io_service::run. Sometimes this run function seems to eat the whole cpu(100%), and sometimes not. I am not very clear about the pattern.
the relevant code:
class Asio {
public:
Asio() :
io_service_(new boost::asio::io_service),
run_() {}
void Start() {
if (!asio_thread_.joinable()) {
run_ = true;
asio_thread_ = std::thread([=] {
Run();
});
}
}
boost::asio::io_service* io_service() { return io_service_.get(); }
protected:
void Run() {
for (;;) {
boost::system::error_code ec;
io_service()->run(ec);
if (run_) {
io_service()->reset();
std::this_thread::sleep_for(std::chrono::milliseconds(100));
} else {
break;
}
}
}
protected:
std::unique_ptr<boost::asio::io_service> io_service_;
std::thread asio_thread_;
std::atomic<bool> run_;
};
When the run function runs normally, below is the callstack
#0 0x00000035f74e9163 in epoll_wait () from /lib64/libc.so.6
#1 0x0000000000b3f6ef in boost::asio::detail::epoll_reactor::run(bool, boost::asio::detail::op_queue<boost::asio::detail::task_io_service_operation>&) ()
#2 0x0000000000b40111 in boost::asio::detail::task_io_service::do_run_one(boost::asio::detail::scoped_lock<boost::asio::detail::posix_mutex>&, boost::asio::detail::task_io_service_thread_info&, boost::system::error_code const&) ()
#3 0x0000000000b3feaf in boost::asio::detail::task_io_service::run(boost::system::error_code&) ()
#4 0x0000000000b403fd in boost::asio::io_service::run(boost::system::error_code&) ()
#5 0x0000000000b3ddc1 in Asio::Run() ()
When the run function behaves abnormally, below is the callstack:
#0 0x00000031bbee53c9 in syscall () from /lib64/libc.so.6
#1 0x00007f831d1d3d68 in std::chrono::_V2::steady_clock::now() () from /usr/local/gcc48/lib64/libstdc++.so.6
#2 0x0000000000b45b6d in boost::asio::detail::chrono_time_traits<std::chrono::_V2::steady_clock, boost::asio::wait_traits<std::chrono::_V2::steady_clock> >::now() ()
#3 0x0000000000b45608 in boost::asio::detail::timer_queue<boost::asio::detail::chrono_time_traits<std::chrono::_V2::steady_clock, boost::asio::wait_traits<std::chrono::_V2::steady_clock> > >::get_ready_timers(boost::asio::detail::op_queue<boost::asio::detail::task_io_service_operation>&) ()
#4 0x0000000000b3f5d7 in boost::asio::detail::timer_queue_set::get_ready_timers(boost::asio::detail::op_queue<boost::asio::detail::task_io_service_operation>&) ()
#5 0x0000000000b3f815 in boost::asio::detail::epoll_reactor::run(bool, boost::asio::detail::op_queue<boost::asio::detail::task_io_service_operation>&) ()
#6 0x0000000000b40111 in boost::asio::detail::task_io_service::do_run_one(boost::asio::detail::scoped_lock<boost::asio::detail::posix_mutex>&, boost::asio::detail::task_io_service_thread_info&, boost::system::error_code const&) ()
#7 0x0000000000b3feaf in boost::asio::detail::task_io_service::run(boost::system::error_code&) ()
#8 0x0000000000b403fd in boost::asio::io_service::run(boost::system::error_code&) ()
#9 0x0000000000b3ddc1 in Asio::Run() ()
In both cases, there're some pending handlers in the io_service, so the io_service::run should not return and should be wait for the event to happen.
Any advise is welcome.
I did further check, it's seems it's due to the boost::asio::steady_timer used. The usage of steady_timer involves the usage of the following pattern:
boost::asio::steady_timer timer;
timer.expires_at(some_expiry, error_code)
timer.async_wait((=)(boost::system::error_code ec) {
// some operation
timer.expires_at(new_expiry, error_code);
timer.asyn_wait(...);
});
Where the timer is wrapped in a shared pointer, and it's safe to copy into the lamda function.