process scheduling and cpuset

[Dmitry Sivachenko]

> On 13 сент. 2015 г., at 19:40, Slawa Olhovchenkov <slw at zxy.spb.ru> wrote:
> 
> On Sun, Sep 13, 2015 at 04:44:40PM +0300, Dmitry Sivachenko wrote:
> 
>> 
>>> On 13 сент. 2015 г., at 16:09, Slawa Olhovchenkov <slw at zxy.spb.ru> wrote:
>>> 
>>> On Sun, Sep 13, 2015 at 02:52:08PM +0300, Dmitry Sivachenko wrote:
>>> 
>>>> Hello,
>>>> 
>>>> I have 32 processor machine (2x CPU E5-2650) running several CPU-bound processes (ULE scheduler).
>>>> 3 processes are 32-threaded, and 8 are single threaded.
>>>> 
>>>> I bind all 3 32-threaded processes to CPUs 0-24 (cpuset -C -l 0-24 -p XXX).
>>>> 
>>>> I expect that the remaining 8 single-threaded processes will (mostly) run on the remaining 25-31 CPU cores and use (almost) 100% cpu each.
>>>> 
>>>> But this is not the case (according to top(1)):  they spend a lot of time on 0-24 CPUs and CPU Idle time is about 10%.
>>>> 
>>>> These are all purely computational programs, in idle system single-threaded programs steadily consume 100% of a core, and 32-threaded programs consume all 32 cores and idle time is zero.
>>>> 
>>>> Is it an ULE scheduler feature or am I doing something wrong?
>>>> 
>>>> The goal is to give a single-threaded program a chance to run when somebody started several 32-threaded processes.
>>> 
>>> You don't have 32 processor machine, you have only 16 processor
>>> machine.
>>> SMT/hyperthreading don't give real processor, SMT "CPU" have
>>> unpredicable power and his load depend on load parent CPU.
>>> 
>>> For example, for my case I see such condition (simpliy) on CPU 0 and 1
>>> (SMT of one real core) with rise load:
>>> 
>>> load 0.1  0.1
>>> load 0.2  0.2
>>> load 0.3  0.3
>>> load 0.4  0.4
>>> load 0.45 0.45
>>> load 0.48 0.48
>>> load 1.00 1.00\
>> 
>> 
>> Yes I know about HT.  But how does this explain why I have 10% of CPU idle?
>> 
>> If I explicitly bind my single-threaded processes to the remaining CPU cores (25-32), they start to receive expected 100% of CPU and overall Idle decreases.
>> 
>> I just expect scheduler to do the same for me.
>> 
> 
> Idle is not goal, goal is lessing task executing time.


Thanks for the explanation.

In my example SMT pairs are numbered with sequential numbers, so 0+1 is one SMT group, 2+3 is second SMT group, and so on.

So in 25-32 range there are several real CPU cores which remain idle while processes are fighting for overloaded 0-24.

When I explicitly pin my single-threaded processes to 25-32 range, they start to receive 100% of CPU (and finish faster to be clear).