Re: [RFC PATCH v2 0/5] sched: modular find_busiest_group()

[Vaidyanathan Srinivasan]

* Peter Zijlstra <peterz@xxxxxxxxxxxxx> [2008-10-14 14:09:13]:

> 
> Hi,
> 
> So the basic issue is sched_group::cpu_power should become more dynamic.

Hi Peter,

This is a good idea.  Dynamically increasing cpu power to some groups
will automatically help power savings when we want to consolidate
better to one cpu package when overall system utilisation is very low.  
 
> There are two driving factors:
> - RT time consumption feedback into CFS
> - dynamic per-cpu power manangement like Intel's Dynamic Speed
> Technology (formerly know as Turbo Mode).
> 
> We currently have sched_group::cpu_power to model SMT. We say that
> multiple threads that share a core are not as powerful as two cores.
> 
> Therefore, we move a task when doing so results in more of that power
> being utilized, resulting in preferring to run tasks on full cores
> instead of SMT siblings.
> 
> 
> RT time
> -------
> 
> So the basic issue is that we cannot know how much cpu-time will be
> consumed by RT tasks (we used to relate that to the number of running
> tasks, but that's utter nonsense).
> 
> Therefore the only way is to measure it and assume the near future looks
> like the near past.
> 
> So why is this an issue.. suppose we have 2 cpus, and 1 cpu is consumed
> for 50% by RT tasks, while the other is fully available to regular
> tasks.
> 
> In that case we'd want to load-balance such that the cpu affected by the
> RT task(s) gets half the load the other cpu has.
> 
> [ I tried modelling this by scaling the load of cpus up, but that fails
> to handle certain cases - for instance 100% RT gets real funny, and it
> fails to properly skip the RT-loaded cores in the low-load situation ]
> 
> Dynamic Speed Technology
> ------------------------
> 
> With cpus actively fiddling with their processing capacity we get into
> similar issues. Again we can measure this, but this would require the
> addition of a clock that measures work instead of time.
> 
> Having that, we can even acturately measure the old SMT case, which has
> always been approximated by a static percentage - even though the actual
> gain is very workload dependent.
> 
> The idea is to introduce sched_work_clock() so that:
> 
>         work_delta / time_delta gives the power for a cpu. <1 means we
>         did less work than a dedicated pipeline, >1 means we did more.

The challenge here is measurement of 'work'.  What will be the
parameter that will be fair for most workloads and easy to measure on
most systems?

* Instructions completion count
* APERF or similar CPU specific counter on x86
* POWER has PURR and SPURR to have a measure of relative work done  


> So, if for example our core's canonical freq would be 2.0GHz but we get
> boosted to 2.2GHz while the other core would get lowered to 1.8GHz we
> can observe and attribute this asymetric power balance.
> 
> [ This assumes that the total power is preserved for non-idle situations
> - is that true?, if not this gets real interesting ]

I would assume total compute power will be preserved over a long
period of time.  But certain workloads can benefit more from acceleration
on the same system challenging the above assumption.

> Also, an SMT thread, when sharing the core with its sibling will get <1,
> but together they might be >1.

In this case what is the normalised value of '1'  It is difficult to
estimate the nominal cpu power with threads.  If we can assume
normalised value to be theoretical max, then sum of both threads can
be less than 1 and will never achieve 1 in practice :)

> Funny corner cases
> ------------------
> 
> Like mentioned in the RT time note, there is the possiblity that a core
> has 0 power (left) for SCHED_OTHER. This has a consequence for the
> balance cpu. Currently we let the first cpu in the domain do the
> balancing, however if that CPU has 0 power it might not be the best
> choice (esp since part of the balancing can be done from softirq context
> - which would basically starve that group).

Agreed, but relative easy to solve compared to other challenges :)
 
> Sched domains
> -------------
> 
> There seems to be a use-case where we need both the cache and the
> package levels. So I wanted to have both levels in there.
> 
> Currently each domain level can only be one of:
> 
> SD_LV_NONE = 0,
> SD_LV_SIBLING,
> SD_LV_MC,
> SD_LV_CPU,
> SD_LV_NODE,
> SD_LV_ALLNODES,
> 
> So to avoid a double domain with 'native' multi-core chips where the
> cache and package level have the same span, I want to encode this
> information in the sched_domain::flags as bits, which means a level can
> be both cache and package.

This will help power savings balance and make the implementation
clean.  You have suggested this previously also.
Similarly collapse the NODE level if it is redundant?

> Over balancing
> --------------
> 
> Lastly, we might need to introduce SD_OVER_BALANCE, which toggles the
> over-balance logic. While over-balancing brings better fairness for a
> number of cases, its also hard on power savings.

I did not understand this over balance.  Can you please explain.

Thanks,
Vaidy

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/