Implementation Guidelines

$ Production

# Tuned

* Engine

Must use the standard gRPC library for the language with default configuration. No custom protobuf serialization.

May optimize gRPC channel settings, thread pools, and use custom protobuf serialization.

No specific rules. Ranked separately from frameworks.

Measures unary gRPC call throughput over cleartext HTTP/2 (h2c). The server implements a simple GetSum RPC that adds two integers.

Connections: 256, 1,024 Concurrent streams per connection: 100

Proto definition

syntax = "proto3";
package benchmark;

service BenchmarkService {
  rpc GetSum (SumRequest) returns (SumReply);
}

message SumRequest {
  int32 a = 1;
  int32 b = 2;
}

message SumReply {
  int32 result = 1;
}

Workload

POST /benchmark.BenchmarkService/GetSum sent as a gRPC unary call over h2c. The load generator (h2load) sends pre-encoded protobuf frames with gRPC headers (content-type: application/grpc, te: trailers).

The request payload is a 9-byte gRPC frame: 5-byte frame header + 4-byte protobuf-encoded SumRequest{a=1, b=2}.

What it measures

gRPC/HTTP2 frame processing overhead
Protocol Buffers serialization/deserialization performance
HTTP/2 stream multiplexing efficiency for gRPC
Server-side gRPC service dispatch latency

Expected request/response

POST /benchmark.BenchmarkService/GetSum HTTP/2
content-type: application/grpc
te: trailers

<binary: 00 00000004 08011002>

HTTP/2 200 OK
content-type: application/grpc

<binary: 00 00000002 0803>

grpc-status: 0

How it differs from HTTP/2 baseline

	Baseline (HTTP/2)	Unary (gRPC)
Protocol	HTTP/2 over TLS	gRPC over h2c
Serialization	Plain text	Protocol Buffers
Port	8443	8080
Request method	GET	POST
Load generator	h2load	h2load (raw gRPC frames)

Parameters

Parameter	Value
RPC	`BenchmarkService/GetSum`
Connections	256, 1,024
Streams per connection	100 (`-m 100`)
Duration	5s
Runs	3 (best taken)
Load generator	h2load

Validation