On Wed, Feb 24, 2021 at 11:29:04PM -0800, Nadav Amit wrote:
> From: Nadav Amit <[email protected]>
>
> Just as applications can use prefetch instructions to overlap
> computations and memory accesses, applications may want to overlap the
> page-faults and compute or overlap the I/O accesses that are required
> for page-faults of different pages.
>
> Applications can use multiple threads and cores for this matter, by
> running one thread that prefetches the data (i.e., faults in the data)
> and another that does the compute, but this scheme is inefficient. Using
> mincore() can tell whether a page is mapped, but might not tell whether
> the page is in the page-cache and does not fault in the data.
>
> Introduce prefetch_page() vDSO-call to prefetch, i.e. fault-in memory
> asynchronously. The semantic of this call is: try to prefetch a page of
> in a given address and return zero if the page is accessible following
> the call. Start I/O operations to retrieve the page if such operations
> are required and there is no high memory pressure that might introduce
> slowdowns.
>
> Note that as usual the page might be paged-out at any point and
> therefore, similarly to mincore(), there is no guarantee that the page
> will be present at the time that the user application uses the data that
> resides on the page. Nevertheless, it is expected that in the vast
> majority of the cases this would not happen, since prefetch_page()
> accesses the page and therefore sets the PTE access-bit (if it is
> clear).
>
> The implementation is as follows. The vDSO code accesses the data,
> triggering a page-fault it is not present. The handler detects based on
> the instruction pointer that this is an asynchronous-#PF, using the
> recently introduce vDSO exception tables. If the page can be brought
> without waiting (e.g., the page is already in the page-cache), the
> kernel handles the fault and returns success (zero). If there is memory
> pressure that prevents the proper handling of the fault (i.e., requires
> heavy-weight reclamation) it returns a failure. Otherwise, it starts an
> I/O to bring the page and returns failure.
>
> Compilers can be extended to issue the prefetch_page() calls when
> needed.
Interesting, but given we've been removing explicit prefetch from some
parts of the kernel how useful is this in actual use? I'm thinking there
should at least be a real user and performance numbers with this before
merging.
> On Feb 25, 2021, at 12:40 AM, Peter Zijlstra <[email protected]> wrote:
>
> On Wed, Feb 24, 2021 at 11:29:04PM -0800, Nadav Amit wrote:
>> From: Nadav Amit <[email protected]>
>>
>> Just as applications can use prefetch instructions to overlap
>> computations and memory accesses, applications may want to overlap the
>> page-faults and compute or overlap the I/O accesses that are required
>> for page-faults of different pages.
>>
>> Applications can use multiple threads and cores for this matter, by
>> running one thread that prefetches the data (i.e., faults in the data)
>> and another that does the compute, but this scheme is inefficient. Using
>> mincore() can tell whether a page is mapped, but might not tell whether
>> the page is in the page-cache and does not fault in the data.
>>
>> Introduce prefetch_page() vDSO-call to prefetch, i.e. fault-in memory
>> asynchronously. The semantic of this call is: try to prefetch a page of
>> in a given address and return zero if the page is accessible following
>> the call. Start I/O operations to retrieve the page if such operations
>> are required and there is no high memory pressure that might introduce
>> slowdowns.
>>
>> Note that as usual the page might be paged-out at any point and
>> therefore, similarly to mincore(), there is no guarantee that the page
>> will be present at the time that the user application uses the data that
>> resides on the page. Nevertheless, it is expected that in the vast
>> majority of the cases this would not happen, since prefetch_page()
>> accesses the page and therefore sets the PTE access-bit (if it is
>> clear).
>>
>> The implementation is as follows. The vDSO code accesses the data,
>> triggering a page-fault it is not present. The handler detects based on
>> the instruction pointer that this is an asynchronous-#PF, using the
>> recently introduce vDSO exception tables. If the page can be brought
>> without waiting (e.g., the page is already in the page-cache), the
>> kernel handles the fault and returns success (zero). If there is memory
>> pressure that prevents the proper handling of the fault (i.e., requires
>> heavy-weight reclamation) it returns a failure. Otherwise, it starts an
>> I/O to bring the page and returns failure.
>>
>> Compilers can be extended to issue the prefetch_page() calls when
>> needed.
>
> Interesting, but given we've been removing explicit prefetch from some
> parts of the kernel how useful is this in actual use? I'm thinking there
> should at least be a real user and performance numbers with this before
> merging.
Can you give me a reference to the “removing explicit prefetch from some
parts of the kernel”?
I will work on an llvm/gcc plugin to provide some performance numbers.
I wanted to make sure that the idea is not a complete obscenity first.
> On Feb 25, 2021, at 12:52 AM, Nadav Amit <[email protected]> wrote:
>
>
>
>> On Feb 25, 2021, at 12:40 AM, Peter Zijlstra <[email protected]> wrote:
>>
>> On Wed, Feb 24, 2021 at 11:29:04PM -0800, Nadav Amit wrote:
>>> From: Nadav Amit <[email protected]>
>>>
>>> Just as applications can use prefetch instructions to overlap
>>> computations and memory accesses, applications may want to overlap the
>>> page-faults and compute or overlap the I/O accesses that are required
>>> for page-faults of different pages.
[
[ snip ]
>> Interesting, but given we've been removing explicit prefetch from some
>> parts of the kernel how useful is this in actual use? I'm thinking there
>> should at least be a real user and performance numbers with this before
>> merging.
>
> Can you give me a reference to the “removing explicit prefetch from some
> parts of the kernel”?
Oh. I get it - you mean we remove we remove the use of explicit memory
prefetch from the kernel code. Well, I don’t think it is really related,
but yes, performance numbers are needed.
On Thu, Feb 25, 2021 at 01:32:56AM -0800, Nadav Amit wrote:
> > On Feb 25, 2021, at 12:52 AM, Nadav Amit <[email protected]> wrote:
> > Can you give me a reference to the “removing explicit prefetch from some
> > parts of the kernel”?
75d65a425c01 ("hlist: remove software prefetching in hlist iterators")
e66eed651fd1 ("list: remove prefetching from regular list iterators")
> Oh. I get it - you mean we remove we remove the use of explicit memory
> prefetch from the kernel code. Well, I don’t think it is really related,
> but yes, performance numbers are needed.
Right, so my main worry was that use of the prefetch instruction
actually hurt performance once the hardware prefetchers got smart
enough, this being a very similar construct (just on a different level
of the stack) should be careful not to suffer the same fate.