SCHED_ULE bug (was Re: cpuminer mines only on one core regardless of "--threads" option)

[Daniel Kalchev]

On 17.01.14 09:47, Andrey Chernov wrote:
> On 16.01.2014 11:20, Alexander wrote:
>>
>> 14.01.2014, 22:32, "Alexander" <ags18 at yandex.ru>:
>>> 14.01.2014, 20:14, "Subbsd" <subbsd at gmail.com>:
>>>
>>>>   On Tue, Jan 14, 2014 at 7:43 PM, Andrey Chernov <ache at freebsd.org> wrote:
>>>>>    On 14.01.2014 17:01, Alexander wrote:
>>>>>>    on Freebsd 9.2 x64 on 5 different PCs I installed net-p2p/cpuminer
>>>>>>    from ports and by "pkg install" - result is the same - minerd mines
>>>>>>    only on one core regardless of "--threads" option.
>>>>>    ...
>>>>>>    # top -P
>>>>>>    CPU 0:  0.0% user,  0.0% nice,  0.0% system,  0.0% interrupt,  100% idle
>>>>>>    CPU 1:  0.0% user,  0.0% nice,  0.0% system,  0.0% interrupt,  100% idle
>>>>>>    CPU 2:  0.0% user,  0.0% nice,  0.0% system,  0.0% interrupt,  100% idle
>>>>>>    CPU 3:  0.0% user,  100% nice,  0.0% system,  0.0% interrupt,  0.0% idle
>>>>>    This is SCHED_ULE bug, I see _all_ processes (not minerd only in
>>>>>    particular) stuck to the last CPU too (top's 'C' column is equal to the
>>>>>    last CPU and never changes), latest -stable i386. It disappears for me
>>>>>    switching to SCHED_4BSD.
>>>>>
>>> reboot didn't help (try on 3 different PCs)
>> ...
>> minerd was started from cron after reboot immediately and broke SMP, so when I disable minerd from cron and reboot - SMP works fine.
>> I think Andrey Chernov said here the reason http://www.freebsd.org/cgi/query-pr.cgi?pr=kern/163585#reply3
>> So this bug is alive - on 9.1 and 9.2 x64 releases.
>> I didn't try SCHED_4BSD yet, cause don't wont to use not-generic kernel because of mining only, but for interest I'll try it on one machine on this weekends.
> I just try it on latest x64 -current, result is exact the same: the bug
> exist with SCHED_ULE and not with SCHED_4BSD. Something is very wrong in
> SCHED_ULE+cpuset_setaffinity chain.
>

This is all very strange. I have  net-p2p/cpuminer compiled on/around 
Dec 6 (on semi-current 9-stable) on several machines, all with SCHED_ULE 
and it has always attached to every single core available.  It also 
respects --threads and attaches to as many cores as I tell it to. All 
this on FreeBSD 9-stable x64. Some machines are with 32 cores.

Daniel