Boxing interface usage is allocating unnecessary heap data

scenario: using the async API but fully sync in reality after checking completion; each call issues an InternalFasterSession boxed on the heap;

You can see this by running a memory profiler on this branch (branch: faster-boxing; project: /test/Benchmarks) in DEBUG mode (release mode runs BDN; you don't want that). The test runs 25,000 operations in a loop; the only heap alloc is:

Type Allocations
| - FASTER.core.ClientSession<FASTER.core.SpanByte, FASTER.core.SpanByte, Input, Output, FASTER.core.Empty, CacheFunctions>.InternalFasterSession 25,000

Since InternalFasterSession is a struct, this is unlikely to be intentional.

This is almost certainly where InternalFasterSession is being passed as a IFasterSession<Key, Value, Input, Output, Context> - for example DoSlowOperation, AsyncOperationInternal, etc, which will force it to be boxed. To avoid boxing, the type must be either:

• typed as itself, InternalFasterSession
• "constrained", i.e. passed (and used as) a generic TSession where TSession : IFasterSession<Key, Value, Input, Output, Context> (used in a few places, such as FasterSession in RmwAsync)

but: any time it is assigned / passed as IFasterSession<...>: it will box, i.e. allocate

Start with the test/Benchmarks project on the faster-boxing branch and run the scenario in DEBUG mode under a memory profiler. Trace InternalFasterSession through DoSlowOperation, AsyncOperationInternal, and the constrained FasterSession usage in RmwAsync, looking for interface-typed paths. Done means the 25,000-operation test no longer allocates an InternalFasterSession per call.