On 12/3/2018 11:56 AM, Valentin Schneider wrote:
> Hi Steve,
> On 26/11/2018 19:06, Steven Sistare wrote:
>> [...]
>>> Mmm I was thinking we could abuse the wrap() and start at
>>> (fls(prev_span) + 1), but we're not guaranteed to have contiguous spans -
>>> the Arm Juno for instance has [0, 3, 4], [1, 2] as MC-level domains, so
>>> that goes down the drain.
>>>
>>> Another thing that has been trotting in my head would be some helper to
>>> create a cpumask from a sparsemask (some sort of sparsemask_span()),
>>> which would let us use the standard mask operators:
>>>
>>> ----->8-----
>>> struct cpumask *overload_span = sparsemask_span(overload_cpus)
>>>
>>> for_each_domain(this_cpu, sd)
>>> for_each_cpu_and(src_cpu, overload_span, sched_domain_span(sd))
>>> <steal_from here>
>>> -----8>-----
>>>
>>> The cpumask could be part of the sparsemask struct to save us the
>>> allocation, and only updated when calling sparsemask_span().
>>
>> I thought of providing something like this along with other sparsemask
>> utility functions, but I decided to be minimalist, and let others add
>> more functions if/when they become needed.
>> this_cpu_cpumask_var_ptr(select_idle_mask)
>> is a temporary that could be used as the destination of the conversion.
>>
>> Also, conversion adds cost, particularly on larger systems. When comparing a
>> cpumask and a sparsemask, it is more efficient to iterate over the smaller
>> set, and test for membership in the larger, such as in try_steal:
>>
>> for_each_cpu(src_cpu, cpu_smt_mask(dst_cpu)) {
>> if (sparsemask_test_elem(src_cpu, overload_cpus)
>>
>>>> To extend stealing across LLC, I would like to keep the per-LLC
>>>> sparsemask,
>>>> but add to each SD a list of sparsemask pointers. The list nodes would be
>>>> private, but the sparsemask structs would be shared. Each list would
>>>> include
>>>> the masks that overlap the SD's members. The list would be a singleton at
>>>> the
>>>> core and LLC levels (same as the socket level for most processors), and
>>>> would
>>>> have multiple elements at the NUMA level.
>>>
>>> I see. As for misfit, creating asym_cpucapacity siblings of the sd_llc_*()
>>> functions seems a bit much - there'd be a lot of redundancy for basically
>>> just a single shared sparsemask, which is why I was rambling about moving
>>> things to root_domain.
>>>
>>> Having different locations where sparsemasks are stored is a bit of a
>>> pain which I'd like to avoid, but if it can't be unified I suppose we'll
>>> have to live with it.
>>
>> I don't follow. A per-LLC sparsemask representing misfits can be allocated
>> with
>> one more line in sd_llc_alloc, and you can steal across LLC using the list I
>> briefly described above.
>
> Ah yes, that would work. Thing is, I had excluded having the misfit masks
> being in the sd_llc_shareds, since from a logical standpoint they don't
> really belong there.
>
> With asymmetric CPU capacities we kind of disregard the cache landscape
Sure, but adding awareness of the cache hierarchy can only make it better,
and a per-LLC mask organization can serve both the overloaded and misfit
use cases quite naturally.
> and focus on, well, CPU asymmetries. There's a few commits laying around
> that forgo some cache usage optimisations for asymmetric systems -
> this one comes to mind:
>
> 9c63e84db29b ("sched/core: Disable SD_PREFER_SIBLING on asymmetric CPU
> capacity domains")
>
> So in truth I was envisioning separate SD_ASYM_CPUCAPACITY-based
> sparsemasks, which is why I was rambling about SD_ASYM_CPUCAPACITY siblings
> of sd_llc_*()... *But* after I had a go at it, it looked to me like that
> was a lot of duplicated code.
I would be happy to review your code and make suggestions to reduce duplication,
and happy to continue to discuss clean and optimal handling for misfits.
However,
I have a request: can we push my patches across the finish line first?
Stealing
for misfits can be its own patch series. Please consider sending your
reviewed-by
for the next version of my series. I will send it soon.
> My root_domain suggestion stems from the fact that we only really need one
> single sparsemask for misfit stealing, and it provides a unique location
> to store the sparsemasks (and you mask them however you want when it comes
> to using them).
>
> Sadly I think that doesn't work as well for cfs_overload_cpus since you
> can't split a sparsemask's chunks over several NUMA nodes, so we'd be
> stuck with an allocation on a single node (but we already do that in some
> places, e.g. for nohz.idle_cpus_mask, so... Is it that bad?).
It can be bad for high memory bandwidth workloads, as the sparsemasks will
be displaced from cache and we incur remote memory latencies on next access.
- Steve
