* why are scheduling domains used in multiprocessor systems
2013-01-09 8:03 why are scheduling domains used in multiprocessor systems Bond
@ 2013-01-10 1:02 ` Peter Teoh
2013-01-10 3:30 ` Preeti U Murthy
1 sibling, 0 replies; 5+ messages in thread
From: Peter Teoh @ 2013-01-10 1:02 UTC (permalink / raw)
To: kernelnewbies
On Wed, Jan 9, 2013 at 4:03 PM, Bond <jamesbond.2k.g@gmail.com> wrote:
> Hi,
> please see this question
>
> http://stackoverflow.com/questions/14229793/what-does-struct-sched-domain-stands-for-in-include-linux-sched-h-scheduling-do
>
> I checked following
> http://lwn.net/Articles/169277/ and following
> http://www.kernel.org/doc/Documentation/scheduler/sched-domains.txt
> the first line of kernel.org doc says
> ............. Each CPU has a "base" scheduling domain (struct
> sched_domain)..............
> and second para says
> " each scheduling domain spans a number of CPUs (stored in the ->span
> field)."
> third para says
> " Each scheduling domain must have one or more CPU
> groups..................
> The intersection of cpumasks from any two of these groups
> MUST be the empty set."
> then some where in doc it says
> "Balancing within a sched domain occurs between groups. That is, each group
> is treated as one entity." the doc in details talks about the
> implementation of
>
> scheduling domains and mentions that CPUs should belong to one of the
> scheduling domain in a way that
> cpumasks intersection should be an empty set
>
> The answer of the question that I want to know is
> why is a scheduling domain actually needed?
>
> _________________________________________
>
> CPU scheduling involving many configuration and factors.
https://www.cs.unm.edu/~eschulte/classes/cs587/data/10.1.1.59.6385.pdf
Goto page 18 for definition of scheduler domain, and it says:
"Each node in a system has a scheduler domain that points to its parent
scheduler domain. A node might be
a uniprocessor system, an SMP system, or a node within a NUMA system."
this complex hierarchies of CPU is normally associated with hardware
physical proximity CPU (just one factors) or the speed of bus that connect
between CPU. Not all CPU are connected to all other CPU, but perhaps only
two or 4 other CPU, and therefore, when u transfer data between CPU, it is
necessary to build these proximities information into the kernel, to
minimize costs of data transfer between CPU.
90% (or more) of supercomputers (with thousands of CPU) are run by Linux
kernel, and clearly each CPU can only have a few neighboring CPU. Other
factors involved power-management: when your processing usage goes down,
u have to shut down the CPU - leaving only the bare minimum to be running.
Organizing in some hierarchies facilitate this scheduling algorithm.
http://www.intel.com/technology/itj/2007/v11i4/9-process/6-linux-scheduler.htm
http://www.cs.stonybrook.edu/~porter/courses/cse506/f12/slides/scheduling.pdf
http://www.cs.stonybrook.edu/~porter/courses/cse506/f12/slides/scheduling2.pdf
--
Regards,
Peter Teoh
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.kernelnewbies.org/pipermail/kernelnewbies/attachments/20130110/7d1577c8/attachment.html
^ permalink raw reply [flat|nested] 5+ messages in thread
* why are scheduling domains used in multiprocessor systems
2013-01-09 8:03 why are scheduling domains used in multiprocessor systems Bond
2013-01-10 1:02 ` Peter Teoh
@ 2013-01-10 3:30 ` Preeti U Murthy
2013-01-10 10:09 ` Bond
1 sibling, 1 reply; 5+ messages in thread
From: Preeti U Murthy @ 2013-01-10 3:30 UTC (permalink / raw)
To: kernelnewbies
On 01/09/2013 01:33 PM, Bond wrote:
> The answer of the question that I want to know is
> why is a scheduling domain actually needed?
Scheduling domains and scheduler groups/cpu groups help to ease the
process of scheduling tasks like:
1.load balancing tasks across cpus.
2.choosing a cpu for a new task to run on.
3.choosing a cpu for a sleeping task to run when it wakes up.
It has a two fold advantage:
1.It organises the cpus in the system very well into groups and hierarchies.
2.It organises the cpus in such a way that it is useful.All cpus which
share an l2 cache belong to one domain.All cpus which share an l3 cache
belong to a higher level domain,which encompasses all the domains which
share the l2 cache.
The advantages that you see with a tree like data structure are similar
here to the advantages of scheduler domains and groups.
Let me explain the use cases with a simple diagram:
_________sd1________
/ \
----------------------
l3 cache
----------------------
--------- ----------
l2 cache l2 cache
--------- ----------
cpu0 cpu1 cpu2 cpu3
\_______/ \________/
sd0 sd0
________sd1_________
/ \
----------------------
l3 cache
----------------------
--------- ----------
l2 cache l2 cache
--------- ----------
cpu4 cpu5 cpu6 cpu7
\_______/ \________/
sd0 sd0
What you see above is a scheduler domain hierarchy.sd1 encompasses sd0s
which happen to be scheduler groups of sd1.Every cpu has a scheduler
domain hierarchy associated with it.For eg.
cpu0->sd=sd0; sd0->parent=sd1.This way through a linked list we can
iterate through all the scheduler domains to which a cpu belongs to.
How does this help?
1.load balancing: Say cpu0 is idle and is ready to pull tasks upon
itself to relieve any other burdened cpu.In the above approach,it first
checks if the other cpus that belong to the first level sched domain
,needs to be relieved of load.Here, cpu1.If so it takes on tasks from
cpu1,else it goes to the higher level domain sd1.If it chooses to
migrate task from cpu1 it is the best thing,because the cache contents
can be utilized;shared cache.no need to fetch from memory again.This is
the first advantage:sched domains are formed based upon the advantages
that hardware has to provide.
If it goes to sd1,then it probes sd1's 'groups',both the sd0s.Here is
the next advantage.It needs information about the sched group alone and
will not bother about the individual cpus in it.it checks if
load(sd0[cpu2,cpu3]) > load(sd0[cpu0,cpu1])
Only if this is true does it go on to see if cpu2/3 is more loaded.If
there were no scheduler domain or groups,we would have to see the states
of cpu2 and cpu3 in two iterations instead of 1 iteration like we are
doing now.
Now scale this problem and solution to 128 cpus! imagine what a mess it
would have been if there was nothing to tell you which cpu would be the
best to relieve load from,in the worst case you would have to iterate
through all the 128 cpus.
But with scheduler domain or groups,say you divide the 128 cpus into
groups of 16 cpus,you would have 8 groups.see which is the busiest,so
that would be 8 iterations,then you would know the busiest group,then
descend down.another 16 iterations.so worst case
8+16 = 24 iterations.And this decrease is only with one level of sched
domain. Imagine if you had more levels,you would make the number of
iterations even lower.
So in short the scheduler domains and groups are a 'divide and conquer
;but conquer as much as possible what is more useful' solution to
scheduling related stuff :)
Thank you
Regards
Preeti U Murthy
^ permalink raw reply [flat|nested] 5+ messages in thread