gRPC Comparison Report — Direct-RID Baseline

This document records the comparison axis and publication state of the new baseline defined by the bench specification. Earlier measurements are not compared with this baseline: the framework row included channel-level node selection and used a different pattern set. The supervisor will fill in new three-run values after measuring on a quiet machine.

1. Comparison Rows

Only the following three rows are compared within one language. Every cell has one target server (B).

Row	Measured path
grpc-<lang>	Standard gRPC unary Echo / Command connected to one server
zlink-<lang>	Raw binding ROUTER↔ROUTER path addressed directly by RID, bypassing the framework
zlink-framework-<lang>	RouteMesh requestToNode / sendToNode addressed directly to the server RID

The framework row no longer measures channel-level round-robin through requestToChannel / sendToChannel. It now has the same 1:1 target-selection condition as gRPC, so node selection is not mixed into the framework/raw ratio.

2. Wire and Patterns

Changing the routing API does not change the application wire. The framework row in all three languages keeps the existing fixed ZLink envelope and the protobuf body BenchPayload { bytes body = 1; }. The first 29 payload bytes remain the measurement header and the rest remains business payload.

Pattern	In-flight meaning	Request throughput basis
request-serial	1	Successful source request/reply completions
request-backpressure	No application ceiling; reached depth is recorded as an outcome	Successful source request/reply completions
send-saturation	8 logical streams, one submission per stream	Active-header messages received by the target

request-window is excluded because forcing depth 100 measures whether a stack reaches that depth rather than isolating per-request cost. The aggregator may still read archived input for diagnosis, but only the three patterns above appear in current tables and judgements.

3. Measurement Contract

The existing contract for warmup, active duration, admission/backpressure, in-flight measurement, post-active drain, error and abandoned accounting, and result JSON fields is unchanged. send-saturation is calculated from target receives, not source submissions. Cells with errors, loss, or a drain-bound violation cannot satisfy publication gates. request-backpressure remains the judgement pattern, evaluated separately at 1024 and 4096 bytes.

4. New Baseline Status

Language	Framework path	Patterns	Baseline values
Java	Direct RID	Three	Round 1 implementation complete; awaiting the new joint baseline run
.NET	Direct RID	Three	Smoke validated; awaiting three-run measurement
C++	Direct RID	Three	Smoke validated; awaiting three-run measurement
Node.js	Direct RID with schema protobuf serializer	Three	Smoke validated; awaiting three-run measurement

Raw-row conversion is separate work in Issues #91–#93. New ratios are published only after both the framework and raw rows satisfy specification §1.3. Short smoke values from this change verify executability only and are not baseline numbers.

5. Result Table Names

Remeasurement output is grouped by pattern and uses these exact row names at each payload size.

Language	gRPC	Raw	Framework
.NET	grpc-dotnet	zlink-dotnet	zlink-framework-dotnet
Java	grpc-java	zlink-java	zlink-framework-java
Node.js	grpc-node	zlink-node	zlink-framework-node
C++	grpc-cpp	zlink-cpp	zlink-framework-cpp

Request patterns use KOPS; send-saturation uses KMSG/s. Ratios computed from the old denominator and old request-window values are not carried into this report.