Numbers

What the benchmark numbers of ruststream-zeromq measure, and what they do not allow you to conclude.

The comparison against a raw client

The three loops of a row are comparable with each other, because they ran in one process, interleaved, over one transport. A row is not comparable with a row another broker crate publishes: there a server sits in the path, and the transports do different work per message.

The peer is a socket in the same process. With no network and no server, the same cost per message is a larger share of the time than it would be between two hosts. Read the percentage as an upper bound on what a deployed service pays.

An overhead marked indistinguishable is smaller than the spread of the rounds of the two loops it compares. A figure below that spread would claim a precision nobody measured, so none is printed.

The cost of the code

An instruction count holds this crate's code, the runtime above it and the zeromq client's framing on the service's thread. The same count costs more time where it misses the cache, so a count is not a time.

Every figure is the steady state, the difference between a run of 1000 messages and one of 2000. Starting the service, accepting the peer's connection and the first allocations behind them are the cold start, paid once.

What they do not mean

A row is one consumer, one queue, a small body and a peer in the same process. It measures what a message costs inside this crate, not what ZeroMQ can carry.

Every loop puts the same three frames on the wire, because that is the layout this crate documents and a foreign peer writes. The header frame goes out empty. A service that sends headers pays this crate's header parse on every message, on top of the figure here.

Nothing settles a message on this transport. The two loops that go through this crate still call ack, which answers Unsupported, and the raw loop has nothing to call. A row another broker publishes carries a real acknowledgement inside its window, and this one carries none.

A pool of worker lanes, several peers and fan-out to many subscribers each answer a different question, and the comparison measures none of them. The code chart counts a batch handler and a reply on DEALER/ROUTER as well.

The numbers are a snapshot of one machine on one day, with the build recorded beside them. They are measured again by hand, on a machine given to the run alone: the differences they show are smaller than the noise of a shared one.