CP013: Affinity asymmetry between execution and memory resource

[hcedwar]

The Affinity execution_context has asymmetry in that it is constructed from an execution resource but supplies a memory resource. Should relationship be between execution resources and memory resources without an intervening execution context?

[AerialMantis]

We decided to adopt a new approach to representing the system topology where there is a separate topology structure for execution resources and for memory resources.

For this we will need to split out the current execution_resource into two classes; for execution resources and for memory resources, and then provide separate ways to query those topologies; for example this_system::execution_resources and this_system::memory_resources. I think this will also mean dropping can_place_agents and can_place_memory as it will be implied that execution resources can place agents and memory resources can place memory.

One concern that we may need to consider is that the term memory resource may conflict with the existing pmr memory resource.

[hcedwar]

On the other hand, std::pmr::memory_resource may be just the right type, with the spec that the object allocates memory conformal with the topology.

struct topological_memory_resource : public std::pmr::memory_resource {
  iterator begin() const noexcept ;
  iterator end() const noexcept ;
  // etc.
};

[AerialMantis]

I was actually thinking that as well that we could used the pmr::memory_resource type as our memory resource type, the issue I ran into was having it be traversable as part of a topology. Though I think your solution here of having a topology type which inherits from pmr::memory_resource but provides a begin and end could work quite nicely.

[mhoemmen]

Sorry I missed Monday's meeting! I'm a bit confused about why the memory resource type itself needs to be iterable. Why wouldn't one just call a function on a memory resource, to get the range of its subordinate resources? Why would the memory resource itself need begin() and end() instance methods?

[mhoemmen]

The reason I ask is that this would, as observed above, require using a subclass of a pmr thing, rather than just the pmr thing.

[mhoemmen]

I get it now -- as long as the resource is a subclass of pmr, it can work as a pmr thing. The subclass resource interface is just for iterating hierarchy.

[mhoemmen]

However, we did bring up an issue today, that since there's no "memory context," the system could unload a driver when the execution context goes away, and thus break extant memory allocations.

[mhoemmen]

Need a lifetime for memory -- e.g., "memory context"

[hcedwar]

A "context" for a memory resource is an "allocator"

[mhoemmen]

@hcedwar I see -- that makes more sense now. This would mean users would create an allocator from a memory resource. Thanks!

See discussion here about execution resource lifetime. It should also be relevant to memory resources, if one replaces "context" with "allocator."