Ulrich Drepper a écrit :
I have never seen performance numbers for this. If it is punishing
existing code in a measurable way I think it's not anacceptable default
behavior.
Here are some numbers. My test program measures the latency of pthread_broadcast
with 1000 pthreads (all threads are blocked on pthread_cond_wait, the time is
measured between the broadcast call and the last woken pthread).
Here are the average latencies after 5000 measures.
[only this patch is used, not the following.
The system is a dual Xeon 2.80GHz with HT enable]
First case: all threads are SCHED_OTHER
* with simple list:
Iterations=5000
Latency (us) min max avg stddev
3869 7400 6656.73 539.35
* with plist:
Iterations=5000
Latency (us) min max avg stddev
3684 7629 6787.97 479.41
Second case: all threads are SCHED_FIFO with priority equally distributed from
priomin to priomax
* with simple list:
Iterations=5000
Latency (us) min max avg stddev
4548 7197 6656.85 463.30
* with plist:
Iterations=5000
Latency (us) min max avg stddev
8289 11752 9720.12 426.45
So, yes it (logically) has a cost, depending of the number of different
priorities used, so it's specially measurable with real-time threads.
With SCHED_OTHER, I suppose that the priorities are not be very distributed.
May be, supposing it makes sense to respect the priority order only for
real-time pthreads, I can register all SCHED_OTHER threads to the same
MAX_RT_PRIO priotity ?
Or do you think this must be set behind a CONFIG* option ?
(Or finally not interesting enough for mainline ?)
--
Pierre
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
