Massive performance loss from OS::sleep hack

[David Xu]

Kris Kennaway wrote:
> Hi,
> 
> I have been running the volano java benchmark 
> (http://www.volano.com/benchmarks.html) on an 8-core i386 system, and 
> out of the box jdk15 on FreeBSD performs extremely poorly.  The system 
> is more than 90% idle, and profiling shows that the ~800 threads in the 
> benchmark are spending most of their time doing short nanosleep() calls.
> 
> 
> I traced it to the following FreeBSD-specific hack in the jdk:
> 
> // XXXBSD: understand meaning and workaround related to yield
> ...
> // XXXBSD: done differently in 1.3.1, take a look
> int os::sleep(Thread* thread, jlong millis, bool interruptible) {
>   assert(thread == Thread::current(),  "thread consistency check");
> ...
> 
>   if (millis <= 0) {
>     // NOTE: workaround for bug 4338139
>     if (thread->is_Java_thread()) {
>       ThreadBlockInVM tbivm((JavaThread*) thread);
> // BSDXXX: Only use pthread_yield here and below if the system thread
> // scheduler gives time slices to lower priority threads when yielding.
> #ifdef __FreeBSD__
>       os_sleep(MinSleepInterval, interruptible);
> #else
>       pthread_yield();
> #endif
> 
> When I removed this hack (i.e. revert to pthread_yield()) I got an 
> immediate 7-fold performance increase, which brings FreeBSD performance 
> on par with Solaris.
> 
> What is the reason why this code is necessary?  Does FreeBSD's 
> sched_yield() really have different semantics to the other operating 
> systems, or was this a libkse bug that was being worked around?
> 
> Kris

Yeah, our sched_yield() really works well, in kernel, if the thread
scheduling policy is SCHED_OTHER (time-sharing), scheduler temporarily
lowers its priority to PRI_MAX_TIMESHARE, it is enough to give some CPU
time to other threads. Why doesn't the UNIX time-sharing work for java ?

Regards,
David Xu