Re: [PATCH] util: NUMA aware memory preallocation

[David Hildenbrand <david@redhat.com>]

On 11.05.22 11:34, Daniel P. Berrangé wrote:
> On Wed, May 11, 2022 at 11:31:23AM +0200, David Hildenbrand wrote:
>>>> Long story short, management application has no way of learning
>>>> TIDs of allocator threads so it can't make them run NUMA aware.
>>>
>>> This feels like the key issue. The preallocation threads are
>>> invisible to libvirt, regardless of whether we're doing coldplug
>>> or hotplug of memory-backends. Indeed the threads are invisible
>>> to all of QEMU, except the memory backend code.
>>>
>>> Conceptually we need 1 or more explicit worker threads, that we
>>> can assign CPU affinity to, and then QEMU can place jobs on them.
>>> I/O threads serve this role, but limited to blockdev work. We
>>> need a generalization of I/O threads, for arbitrary jobs that
>>> QEMU might want to farm out to specific numa nodes.
>>
>> At least the "-object iothread" thingy can already be used for actions
>> outside of blockdev. virtio-balloon uses one for free page hinting.
> 
> Ah that's good to know, so my idea probably isn't so much work as
> I thought it might be.

Looking into the details, iothreads are the wrong tool to use here.

I can imagine use cases where you'd want to perform preallcoation
* Only one some specific CPUs of a NUMA node (especially, not ones where 
  we pinned VCPUs)
* On CPUs that are on a different NUMA node then the actual backend 
  memory ... just thinking about all of the CPU-less nodes for PMEM and 
  CXL that we're starting to see.


So ideally, we'd let the user configure either
* Set of physical CPUs to use (low hanging fruit)
* Set of NUMA nodes to use (harder)
and allow libvirt to easily configure it similarly by pinning threads.

As CPU affinity is inherited when creating a new thread, so here is one
IMHO reasonable simple thing to get the job done and allow for such
flexibility.


Introduce a "thread-context" user-creatable object that is used to
configure CPU affinity.

Internally, thread-context creates exactly one thread called "TC $id".
That thread serves no purpose besides spawning new threads that inherit the
affinity.

Internal users (like preallocation, but we could reuse the same concept for other
non-io threads, such as userfaultfd, and some other potential non-io thread
users) simply use that thread context to spawn new threads.


Spawned threads get called "TC $id/$threadname", whereby $threadname is the
ordinary name supplied by the internal user. This could be used to identify
long-running threads if needed in the future.

It's worth nothing that this is not a thread pool.


a) Ordinary cmdline usage:

-object thread-context,id="tc1",cpus=0-9,cpus=12

QEMU tries setting the CPU affinity and fails if that's impossible.

-object memory-backend-file,...,prealloc=on,prealloc-threads=16,prealloc-thread-context=tc1

When a context is set, preallcoation code will use the thread-context to spawn threads.


b) Libvirt, ..., usage:

-object thread-context,id="tc1"

Then, libvirt identifies and sets the affinity for the "TC tc1" thread.

-object memory-backend-file,...,prealloc=on,prealloc-threads=16,prealloc-thread-context=tc1



thread-context can be reused for successive preallcoation etc, obviously.


-- 
Thanks,

David / dhildenb