On 14/03/16(Mon) 16:05, Michal Mazurek wrote: > On 04:41:05, 13.03.16, Juan Francisco Cantero Hurtado wrote: > > Here are the commands: > > ... > > ffmpeg > > ... > > Thank you for this. > > ffmpeg runs differently from gcc or make - it creates a lot of threads. > I can verify that it is indeed slower. Instead of spending 2 seconds in > 'system' it takes 30 or 40 seconds on the new scheduler. > > After some tests I realised that ffmpeg, when running on the present BSD > scheduler, will call sys_sched_yield() 321,068 times. When running on the > new scheduler, it will call sys_sched_yield() 2,507,894 times - nearly ten > times that. The bulk of the calls are made from this function: > > lib/librthread/rthread.c: > void > _spinlock(volatile struct _spinlock *lock) > { > while (_atomic_lock(&lock->ticket)) > sched_yield(); > } > > To verify I replaced sched_yield() with usleep(): > > void > _spinlock(volatile struct _spinlock *lock) > { > while (_atomic_lock(&lock->ticket)) > usleep(1); > } > > The number of calls to yield() dropped to 4,576.
This is really similar to what I observed with Firefox and Chrome. > This is where I get stuck, I don't know how to replace the call to > sched_yield(), or whether it is a good idea to do so. Any advice? I'd assume the problem is not in the _spinlock() implementation itself but rather on the code calling this lock. Do you know which code is triggering this? See r1.13 of lib/librthread/rthread_libc.c, this is an example where a the use of a spinlock created similar symptoms.