Add an out-of-process memory benchmark for active, stalled and churning connections

.gitignore

@@ -2,6 +2,7 @@ target/
 node_modules/
 dist/
 dist-test/
+bench-results/
 perf.data*
 *.node
 Cargo.lock

Cargo.toml

@@ -27,6 +27,13 @@ ring        = "0.17"
 md-5        = "0.10"
 sha2        = "0.10"
 zeroize     = "1"
+# `disable_initial_exec_tls`: jemalloc's default TLS model cannot be dlopen()ed, which is how Node
+# loads this cdylib.
+tikv-jemallocator = { version = "0.6", optional = true, features = ["disable_initial_exec_tls"] }
+
+[features]
+# Memory-benchmark arm only (see src/lib.rs); never enabled in published builds.
+jemalloc = ["dep:tikv-jemallocator"]
 
 [dev-dependencies]
 # `start_paused` runtimes for the liveness timer tests; not compiled into the cdylib.

README.md

@@ -945,6 +945,9 @@ Two caveats when sizing a node from this:
   rustls session state per connection, plus kernel socket memory (bounded in aggregate by
   `net.ipv4.tcp_mem`). Measure the actual slope — `activeConnections` from the admin endpoint
   against process RSS — before deciding a connection target is reachable.
+  `npm run benchmark:memory-oop` (`__test__/bench-copy-memory-oop.ts`) does this with symphony in
+  its own process under parked, active, slow-reader and churn load; the other `benchmark*` scripts
+  host the proxy inside their own load generator, so their RSS is mostly the client's.
 - Buffer memory is not what caps connection count. File descriptors (2 per proxied connection,
   against `RLIMIT_NOFILE`) and `nf_conntrack_max` bind first; see
   [Linux kernel tuning](#linux-kernel-tuning).