Regarding HMM

Regarding HMM

[Valmiki]

Hi All,

Im trying to understand heterogeneous memory management, i have 
following doubts.

If HMM is being used we dont have to use DMA controller on device for 
memory transfers ?
Without DMA if software is managing page faults and migrations, will 
there be any performance impacts ?

Is HMM targeted for any specific use cases where DMA controller is not 
there on device ?

Regards,
Valmiki

Re: Regarding HMM

[Ralph Campbell]

On 8/18/20 12:15 AM, Valmiki wrote:
> Hi All,
> 
> Im trying to understand heterogeneous memory management, i have following doubts.
> 
> If HMM is being used we dont have to use DMA controller on device for memory transfers ?
> Without DMA if software is managing page faults and migrations, will there be any performance impacts ?
> 
> Is HMM targeted for any specific use cases where DMA controller is not there on device ?
> 
> Regards,
> Valmiki
> 

There are two APIs that are part of "HMM" and are independent of each other.

hmm_range_fault() is for getting the physical address of a system resident memory page that
a device can map but is not pinned in the usual way I/O increases the page reference count
to pin the page. The device driver has to handle invalidation callbacks to remove the device
mapping. This lets the device access the page without moving it.

migrate_vma_setup(), migrate_vma_pages(), and migrate_vma_finalize() are used by the device
driver to migrate data to device private memory. After migration, the system memory is freed
and the CPU page table holds an invalid PTE that points to the device private struct page
(similar to a swap PTE). If the CPU process faults on that address, there is a callback
to the driver to migrate it back to system memory. This is where device DMA engines can
be used to copy data to/from system memory and device private memory.

The use case for the above is to be able to run code such as OpenCL on GPUs and CPUs using
the same virtual addresses without having to call special memory allocators.
In other words, just use mmap() and malloc() and not clSVMAlloc().

There is a performance consideration here. If the GPU accesses the data over PCIe to
system memory, there is much less bandwidth than accessing local GPU memory. If the
data is to be accessed/used many times, it can be more efficient to migrate the data
to local GPU memory. If the data is only accessed a few times, then it is probably
more efficient to map system memory.

Re: Regarding HMM

[John Hubbard]

On 8/18/20 10:06 AM, Ralph Campbell wrote:
> 
> On 8/18/20 12:15 AM, Valmiki wrote:
>> Hi All,
>>
>> Im trying to understand heterogeneous memory management, i have following doubts.
>>
>> If HMM is being used we dont have to use DMA controller on device for memory transfers ?

Hi,

Nothing about HMM either requires or prevents using DMA controllers.

>> Without DMA if software is managing page faults and migrations, will there be any performance 
>> impacts ?
>>
>> Is HMM targeted for any specific use cases where DMA controller is not there on device ?
>>
>> Regards,
>> Valmiki
>>
> 
> There are two APIs that are part of "HMM" and are independent of each other.
> 
> hmm_range_fault() is for getting the physical address of a system resident memory page that
> a device can map but is not pinned in the usual way I/O increases the page reference count
> to pin the page. The device driver has to handle invalidation callbacks to remove the device
> mapping. This lets the device access the page without moving it.
> 
> migrate_vma_setup(), migrate_vma_pages(), and migrate_vma_finalize() are used by the device
> driver to migrate data to device private memory. After migration, the system memory is freed
> and the CPU page table holds an invalid PTE that points to the device private struct page
> (similar to a swap PTE). If the CPU process faults on that address, there is a callback
> to the driver to migrate it back to system memory. This is where device DMA engines can
> be used to copy data to/from system memory and device private memory.
> 
> The use case for the above is to be able to run code such as OpenCL on GPUs and CPUs using
> the same virtual addresses without having to call special memory allocators.
> In other words, just use mmap() and malloc() and not clSVMAlloc().
> 
> There is a performance consideration here. If the GPU accesses the data over PCIe to
> system memory, there is much less bandwidth than accessing local GPU memory. If the
> data is to be accessed/used many times, it can be more efficient to migrate the data
> to local GPU memory. If the data is only accessed a few times, then it is probably
> more efficient to map system memory.
> 

Ralph, that's a good write-up!

Valmiki, did you already read Documentation/vm/hmm.rst, before posting your question?

It's OK to say "no"--I'm not asking in order to criticize, but in order to calibrate
the documentation. Because, we should consider merging in Ralph's write-up above
into hmm.rst, depending on if it helps (which I expect it does, but I'm tainted by
reading hmm.rst too many times and now I can't see what might be missing).

Any time someone new tries to understand the system, it's an opportunity to "unit test"
the documentation. Ideally, hmm.rst would answer many of a first-time reader's questions,
that's where we'd like to end up.


thanks,
-- 
John Hubbard
NVIDIA

Re: Regarding HMM

[Valmiki]

On 19-08-2020 02:05 am, John Hubbard wrote:
> On 8/18/20 10:06 AM, Ralph Campbell wrote:
>>
>> On 8/18/20 12:15 AM, Valmiki wrote:
>>> Hi All,
>>>
>>> Im trying to understand heterogeneous memory management, i have 
>>> following doubts.
>>>
>>> If HMM is being used we dont have to use DMA controller on device for 
>>> memory transfers ?
> 
> Hi,
> 
> Nothing about HMM either requires or prevents using DMA controllers.
> 
>>> Without DMA if software is managing page faults and migrations, will 
>>> there be any performance impacts ?
>>>
>>> Is HMM targeted for any specific use cases where DMA controller is 
>>> not there on device ?
>>>
>>> Regards,
>>> Valmiki
>>>
>>
>> There are two APIs that are part of "HMM" and are independent of each 
>> other.
>>
>> hmm_range_fault() is for getting the physical address of a system 
>> resident memory page that
>> a device can map but is not pinned in the usual way I/O increases the 
>> page reference count
>> to pin the page. The device driver has to handle invalidation 
>> callbacks to remove the device
>> mapping. This lets the device access the page without moving it.
>>
>> migrate_vma_setup(), migrate_vma_pages(), and migrate_vma_finalize() 
>> are used by the device
>> driver to migrate data to device private memory. After migration, the 
>> system memory is freed
>> and the CPU page table holds an invalid PTE that points to the device 
>> private struct page
>> (similar to a swap PTE). If the CPU process faults on that address, 
>> there is a callback
>> to the driver to migrate it back to system memory. This is where 
>> device DMA engines can
>> be used to copy data to/from system memory and device private memory.
>>
>> The use case for the above is to be able to run code such as OpenCL on 
>> GPUs and CPUs using
>> the same virtual addresses without having to call special memory 
>> allocators.
>> In other words, just use mmap() and malloc() and not clSVMAlloc().
>>
>> There is a performance consideration here. If the GPU accesses the 
>> data over PCIe to
>> system memory, there is much less bandwidth than accessing local GPU 
>> memory. If the
>> data is to be accessed/used many times, it can be more efficient to 
>> migrate the data
>> to local GPU memory. If the data is only accessed a few times, then it 
>> is probably
>> more efficient to map system memory.
>>
> 
> Ralph, that's a good write-up!
> 
> Valmiki, did you already read Documentation/vm/hmm.rst, before posting 
> your question?
> 
> It's OK to say "no"--I'm not asking in order to criticize, but in order 
> to calibrate
> the documentation. Because, we should consider merging in Ralph's 
> write-up above
> into hmm.rst, depending on if it helps (which I expect it does, but I'm 
> tainted by
> reading hmm.rst too many times and now I can't see what might be missing).
> 
> Any time someone new tries to understand the system, it's an opportunity 
> to "unit test"
> the documentation. Ideally, hmm.rst would answer many of a first-time 
> reader's questions,
> that's where we'd like to end up.
> 
> 
> thanks,
Hi John, i did give initial reading, but Ralph details on migration_* 
API's above helped a lot to clarify things better, yes adding above 
details into hmm.rst would help beginners.

Regards,
Valmiki

Re: Regarding HMM

[Valmiki]

On 18-08-2020 10:36 pm, Ralph Campbell wrote:
> 
> On 8/18/20 12:15 AM, Valmiki wrote:
>> Hi All,
>>
>> Im trying to understand heterogeneous memory management, i have 
>> following doubts.
>>
>> If HMM is being used we dont have to use DMA controller on device for 
>> memory transfers ?
>> Without DMA if software is managing page faults and migrations, will 
>> there be any performance impacts ?
>>
>> Is HMM targeted for any specific use cases where DMA controller is not 
>> there on device ?
>>
>> Regards,
>> Valmiki
>>
> 
> There are two APIs that are part of "HMM" and are independent of each 
> other.
> 
> hmm_range_fault() is for getting the physical address of a system 
> resident memory page that
> a device can map but is not pinned in the usual way I/O increases the 
> page reference count
> to pin the page. The device driver has to handle invalidation callbacks 
> to remove the device
> mapping. This lets the device access the page without moving it.
> 
> migrate_vma_setup(), migrate_vma_pages(), and migrate_vma_finalize() are 
> used by the device
> driver to migrate data to device private memory. After migration, the 
> system memory is freed
> and the CPU page table holds an invalid PTE that points to the device 
> private struct page
> (similar to a swap PTE). If the CPU process faults on that address, 
> there is a callback
> to the driver to migrate it back to system memory. This is where device 
> DMA engines can
> be used to copy data to/from system memory and device private memory.
> 
> The use case for the above is to be able to run code such as OpenCL on 
> GPUs and CPUs using
> the same virtual addresses without having to call special memory 
> allocators.
> In other words, just use mmap() and malloc() and not clSVMAlloc().
> 
> There is a performance consideration here. If the GPU accesses the data 
> over PCIe to
> system memory, there is much less bandwidth than accessing local GPU 
> memory. If the
> data is to be accessed/used many times, it can be more efficient to 
> migrate the data
> to local GPU memory. If the data is only accessed a few times, then it 
> is probably
> more efficient to map system memory.
Thanks Ralph for the clarification.