can the scheduler decide to schedule an interrupted but runnable thread on another CPU core? What are the implications for code?
Andrey Chernov
ache at freebsd.org
Fri Feb 14 22:36:24 UTC 2014
On 15.02.2014 2:08, Andrey Chernov wrote:
> On 14.02.2014 23:10, John Baldwin wrote:
>
>>> Due to this bug, not fixed yet, the real picture is more complex:
>>> http://www.freebsd.org/cgi/query-pr.cgi?pr=kern/163585
>>
>> Eh, that bug report has no useful details, as in, it doesn't list the
>> actual commands run. If you do 'cpuset -l 6 -s 1' to force all
>> processes to only use CPU6, then yes, of course the other CPUs are idle
>> because that's what you _asked_ for. AFAICT, that is all the original
>> reporter did. At work we regularly add and remove CPUs from the
>> default set (set 1) on hundreds of machines every day with ULE without
>> any issues.
>
> Probably original report lack certain commands, but I provide the link
> to the port which reproduces this bug too. All threads there are
> assigned to the _different_ CPUs and appears as result on single one
> with SCHED_ULE (not with SCHED_4BSD). And it is what original reporter
> mean too. It surely happens, maybe not the first time, but on 2nd-3rd.
> It means that cpuset_setaffinity() is completely broken form SCHED_ULE
> at least for 3 years.
>
This is code example from cpuminer port, in case you are interested, it is very simple:
static inline void affine_to_cpu(int id, int cpu)
{
cpuset_t set;
CPU_ZERO(&set);
CPU_SET(cpu, &set);
cpuset_setaffinity(CPU_LEVEL_WHICH, CPU_WHICH_CPUSET, -1, sizeof(cpuset_t), &set);
}
...
num_processors = sysconf(_SC_NPROCESSORS_CONF);
if (!opt_n_threads)
opt_n_threads = num_processors;
...
In the thread itself:
struct thr_info *mythr = userdata;
int thr_id = mythr->id;
if (num_processors > 1 && opt_n_threads % num_processors == 0) {
if (!opt_quiet)
applog(LOG_INFO, "Binding thread %d to cpu %d",
thr_id, thr_id % num_processors);
affine_to_cpu(thr_id, thr_id % num_processors);
}
--
http://ache.vniz.net/
More information about the freebsd-hackers
mailing list