Small Performance Improvements

[fabianfett]

This PR has a number of smaller perf improvements. In my measuring they account for about a 3% perf improvement with tiny string payloads. The perf improvement should be much larger for larger invocation payloads.

Modifications:

• Replaced HTTPHandler with NIOHTTPClientResponseAggregator, to fix a CoW issue. This will have by far the biggest speed impact. Especially for large payloads.
• Ensured that we only use one EventLoop in tests.
• Uses syncOperations to setup the HTTPClient.

Result:

Faster Swift Lambda Runtime.

[fabianfett]

This is a cleanup change. public access control in a internal struct doesn't make much sense.

[fabianfett]

We have a fairly large performance issue here... Since body's internal storage is referenced from self.readState and from body, we incur a CoW on every append to the local buffer. We could fix this by an easy CoW prevention like:

case .body(let head, var body):
    self.readState = .idle // removes the self.readState reference. Therefore no Cow necessary
    body.writeBuffer(&bodyPart)
    self.readState = .body(head, body)

However since October of last year NIO has a NIOHTTPClientResponseAggregator. I think we should use that instead. All the http write logic from HTTPHandler has been moved into UnaryHandler directly.

[tomerd]

moving to NIOHTTPClientResponseAggregator makes perfect sense. I wonder tho if moving all the extra functionality to UnaryHandler is overloading its intent - to be a unary handler. not sure the performance costs of having another handlers, if low, we can consider splitting that up, alternatively we can rename UnaryHandler to something more appropriate givens expanded role

[fabianfett]

I have removed the keepAlive option here, since neither tested nor used anywhere in the project. The HTTPClient should always prefer keep-alive but should be able to close a connection after a connection: close header. There is a test that ensures that the HTTPClient works correctly when a connection: close header is received.

[tomerd]

I think the intent was to be able and force closed connection by sending that header from the client, but this is fine.

[fabianfett]

This makes Xcodes debug view much easier to deal with! And we only need a single core for the MockServer anyway.

[tomerd]

good improvement here. one dilemma regarding the expanded role of UnaryHandler to decide before we merge

[tomerd]

<3