Re: [PATCH RESEND 1/4] memory tiering: add abstract distance calculation algorithms management

From: Alistair Popple
Date: Tue Aug 22 2023 - 03:19:02 EST



"Huang, Ying" <ying.huang@xxxxxxxxx> writes:

> Alistair Popple <apopple@xxxxxxxxxx> writes:
>
>> "Huang, Ying" <ying.huang@xxxxxxxxx> writes:
>>
>>> Alistair Popple <apopple@xxxxxxxxxx> writes:
>>>
>>>> "Huang, Ying" <ying.huang@xxxxxxxxx> writes:
>>>>
>>>>> Hi, Alistair,
>>>>>
>>>>> Sorry for late response. Just come back from vacation.
>>>>
>>>> Ditto for this response :-)
>>>>
>>>> I see Andrew has taken this into mm-unstable though, so my bad for not
>>>> getting around to following all this up sooner.
>>>>
>>>>> Alistair Popple <apopple@xxxxxxxxxx> writes:
>>>>>
>>>>>> "Huang, Ying" <ying.huang@xxxxxxxxx> writes:
>>>>>>
>>>>>>> Alistair Popple <apopple@xxxxxxxxxx> writes:
>>>>>>>
>>>>>>>> "Huang, Ying" <ying.huang@xxxxxxxxx> writes:
>>>>>>>>
>>>>>>>>> Alistair Popple <apopple@xxxxxxxxxx> writes:
>>>>>>>>>
>>>>>>>>>>>>> While other memory device drivers can use the general notifier chain
>>>>>>>>>>>>> interface at the same time.
>>>>>>>>>>
>>>>>>>>>> How would that work in practice though? The abstract distance as far as
>>>>>>>>>> I can tell doesn't have any meaning other than establishing preferences
>>>>>>>>>> for memory demotion order. Therefore all calculations are relative to
>>>>>>>>>> the rest of the calculations on the system. So if a driver does it's own
>>>>>>>>>> thing how does it choose a sensible distance? IHMO the value here is in
>>>>>>>>>> coordinating all that through a standard interface, whether that is HMAT
>>>>>>>>>> or something else.
>>>>>>>>>
>>>>>>>>> Only if different algorithms follow the same basic principle. For
>>>>>>>>> example, the abstract distance of default DRAM nodes are fixed
>>>>>>>>> (MEMTIER_ADISTANCE_DRAM). The abstract distance of the memory device is
>>>>>>>>> in linear direct proportion to the memory latency and inversely
>>>>>>>>> proportional to the memory bandwidth. Use the memory latency and
>>>>>>>>> bandwidth of default DRAM nodes as base.
>>>>>>>>>
>>>>>>>>> HMAT and CDAT report the raw memory latency and bandwidth. If there are
>>>>>>>>> some other methods to report the raw memory latency and bandwidth, we
>>>>>>>>> can use them too.
>>>>>>>>
>>>>>>>> Argh! So we could address my concerns by having drivers feed
>>>>>>>> latency/bandwidth numbers into a standard calculation algorithm right?
>>>>>>>> Ie. Rather than having drivers calculate abstract distance themselves we
>>>>>>>> have the notifier chains return the raw performance data from which the
>>>>>>>> abstract distance is derived.
>>>>>>>
>>>>>>> Now, memory device drivers only need a general interface to get the
>>>>>>> abstract distance from the NUMA node ID. In the future, if they need
>>>>>>> more interfaces, we can add them. For example, the interface you
>>>>>>> suggested above.
>>>>>>
>>>>>> Huh? Memory device drivers (ie. dax/kmem.c) don't care about abstract
>>>>>> distance, it's a meaningless number. The only reason they care about it
>>>>>> is so they can pass it to alloc_memory_type():
>>>>>>
>>>>>> struct memory_dev_type *alloc_memory_type(int adistance)
>>>>>>
>>>>>> Instead alloc_memory_type() should be taking bandwidth/latency numbers
>>>>>> and the calculation of abstract distance should be done there. That
>>>>>> resovles the issues about how drivers are supposed to devine adistance
>>>>>> and also means that when CDAT is added we don't have to duplicate the
>>>>>> calculation code.
>>>>>
>>>>> In the current design, the abstract distance is the key concept of
>>>>> memory types and memory tiers. And it is used as interface to allocate
>>>>> memory types. This provides more flexibility than some other interfaces
>>>>> (e.g. read/write bandwidth/latency). For example, in current
>>>>> dax/kmem.c, if HMAT isn't available in the system, the default abstract
>>>>> distance: MEMTIER_DEFAULT_DAX_ADISTANCE is used. This is still useful
>>>>> to support some systems now. On a system without HMAT/CDAT, it's
>>>>> possible to calculate abstract distance from ACPI SLIT, although this is
>>>>> quite limited. I'm not sure whether all systems will provide read/write
>>>>> bandwith/latency data for all memory devices.
>>>>>
>>>>> HMAT and CDAT or some other mechanisms may provide the read/write
>>>>> bandwidth/latency data to be used to calculate abstract distance. For
>>>>> them, we can provide a shared implementation in mm/memory-tiers.c to map
>>>>> from read/write bandwith/latency to the abstract distance. Can this
>>>>> solve your concerns about the consistency among algorithms? If so, we
>>>>> can do that when we add the second algorithm that needs that.
>>>>
>>>> I guess it would address my concerns if we did that now. I don't see why
>>>> we need to wait for a second implementation for that though - the whole
>>>> series seems to be built around adding a framework for supporting
>>>> multiple algorithms even though only one exists. So I think we should
>>>> support that fully, or simplfy the whole thing and just assume the only
>>>> thing that exists is HMAT and get rid of the general interface until a
>>>> second algorithm comes along.
>>>
>>> We will need a general interface even for one algorithm implementation.
>>> Because it's not good to make a dax subsystem driver (dax/kmem) to
>>> depend on a ACPI subsystem driver (acpi/hmat). We need some general
>>> interface at subsystem level (memory tier here) between them.
>>
>> I don't understand this argument. For a single algorithm it would be
>> simpler to just define acpi_hmat_calculate_adistance() and a static
>> inline version of it that returns -ENOENT when !CONFIG_ACPI than adding
>> a layer of indirection through notifier blocks. That breaks any
>> dependency on ACPI and there's plenty of precedent for this approach in
>> the kernel already.
>
> ACPI is a subsystem, so it's OK for dax/kmem to depends on CONFIG_ACPI.
> But HMAT is a driver of ACPI subsystem (controlled via
> CONFIG_ACPI_HMAT). It's not good for a driver of DAX subsystem
> (dax/kmem) to depend on a *driver* of ACPI subsystem.
>
> Yes. Technically, there's no hard wall to prevent this. But I think
> that a good design should make drivers depends on subsystems or drivers
> of the same subsystem, NOT drivers of other subsystems.

Thanks, I wasn't really thinking of HMAT as an ACPI driver. I understand
where you're coming from but I really don't see the problem with using a
static inline. It doesn't create dependencies (you could still use
dax/kmem without ACPI) and results in smaller and easier to follow code.

IMHO it's far more obvious that a call to acpi_hmat_calcaulte_adist()
returns either a default if ACPI HMAT isn't configured or a calculated
value than it is to figure out what notifiers may or may not be
registered at runtime and what priority they may be called in from
mt_calc_adistance().

It appears you think that is a bad design, but I don't understand
why. What does this approach give us that a simpler approach wouldn't?