Add H2 connection pool with per-authority multiplexing

crates/service-client/Cargo.toml

@@ -25,26 +25,41 @@ base64 = { workspace = true }
 bs58 = { workspace = true }
 bytes = { workspace = true }
 bytestring = { workspace = true }
+dashmap = { workspace = true }
+derive_builder = { workspace = true }
 futures = { workspace = true }
 h2 = "0.4.12"
 http = { workspace = true }
+http-body = { workspace = true }
 http-body-util = { workspace = true }
 http-serde = { workspace = true }
 hyper = { workspace = true, features = ["http1", "http2", "client"] }
 hyper-rustls = { workspace = true }
 hyper-util = { workspace = true, features = ["client-legacy"] }
 jsonwebtoken = { workspace = true }
+parking_lot = { workspace = true }
 pem = { version = "3.0.6" }
+pin-project = { workspace = true }
 ring = { version = "0.17.14" }
 rustls = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
 serde_with = { workspace = true }
 thiserror = { workspace = true }
-tower-service = { version = "0.3" }
+tokio = { workspace = true }
+tokio-util = { workspace = true }
+tokio-rustls = "0.26"
+tower = { workspace = true }
 tracing = { workspace = true }
 zstd = { workspace = true }
 
 [dev-dependencies]
+criterion = { workspace = true, features = ["async_tokio"] }
 googletest = { workspace = true }
-tempfile = { workspace = true }
+pprof = { version = "0.15", features = ["criterion", "flamegraph"] }
+tempfile = { workspace = true }
+tokio-stream = {workspace = true}
+
+[[bench]]
+name = "h2_pool_benchmark"
+harness = false

crates/service-client/benches/README.md

@@ -0,0 +1,73 @@
+# Service Client Benchmarks
+
+Compares the custom H2 connection pool against hyper_util's legacy client and the raw h2 client using in-memory duplex streams.
+
+## Benchmark groups
+
+| Group | What it measures |
+|-------|-----------------|
+| `sequential` | Single request latency |
+| `concurrent/{10,50}` | Throughput under H2 multiplexing |
+| `body-{1KB,64KB}` | Data echo throughput |
+
+## Running benchmarks
+
+Run all benchmarks:
+
+```bash
+cargo bench -p restate-service-client --bench h2_pool_benchmark
+```
+
+Dry-run (verify they execute without measuring):
+
+```bash
+cargo bench -p restate-service-client --bench h2_pool_benchmark -- --test
+```
+
+Run a single benchmark by name filter:
+
+```bash
+cargo bench -p restate-service-client --bench h2_pool_benchmark -- "sequential/custom-pool"
+```
+
+## CPU profiling with pprof (built-in)
+
+The benchmarks include [pprof](https://github.com/tikv/pprof-rs) integration that generates flamegraph SVGs.
+
+### Prerequisites
+
+On Linux, allow perf events:
+
+```bash
+sudo sysctl kernel.perf_event_paranoid=-1
+```
+
+### Profile a benchmark
+
+Pass `--profile-time=<seconds>` to activate profiling:
+
+```bash
+cargo bench -p restate-service-client --bench h2_pool_benchmark -- "sequential/custom-pool" --profile-time=30
+```
+
+The flamegraph is written to:
+
+```
+target/criterion/sequential/custom-pool/profile/flamegraph.svg
+```
+
+Open it in a browser for an interactive view.
+
+## CPU profiling with samply (external)
+
+[samply](https://github.com/mstange/samply) can profile the benchmark binary without any code changes.
+
+```bash
+# Build the benchmark binary (release mode)
+cargo bench -p restate-service-client --bench h2_pool_benchmark --no-run
+
+# Find and run with samply
+samply record target/release/deps/h2_pool_benchmark-* --bench "sequential/custom-pool" --profile-time=30
+```
+
+This opens the Firefox Profiler UI automatically.