perf: increase end-to-end transfer TPS

.env.devnet

@@ -25,7 +25,11 @@ CONTRA_SIGVERIFY_QUEUE_SIZE=10000000
 CONTRA_SIGVERIFY_WORKERS=32
 CONTRA_WRITE_MAX_CONNECTIONS=1000000
 CONTRA_READ_MAX_CONNECTIONS=100000
+CONTRA_PG_MAX_CONNECTIONS=32
 CONTRA_MAX_TX_PER_BATCH=64
+CONTRA_BATCH_DEADLINE_MS=10
+CONTRA_BATCH_CHANNEL_CAPACITY=16
+CONTRA_MAX_SVM_WORKERS=8
 CONTRA_ADMIN_KEYS=admin_pubkey
 
 # Gateway

.env.example

@@ -24,7 +24,15 @@ CONTRA_SIGVERIFY_QUEUE_SIZE=10000000
 CONTRA_SIGVERIFY_WORKERS=32
 CONTRA_WRITE_MAX_CONNECTIONS=1000000
 CONTRA_READ_MAX_CONNECTIONS=100000
+# Postgres pool size per node (default 32, capped at 256).
+CONTRA_PG_MAX_CONNECTIONS=32
 CONTRA_MAX_TX_PER_BATCH=64
+# Sequencer batch-flush deadline (ms).
+CONTRA_BATCH_DEADLINE_MS=10
+# Sequencer→executor batch channel capacity.
+CONTRA_BATCH_CHANNEL_CAPACITY=16
+# Executor parallel SVM worker cap.
+CONTRA_MAX_SVM_WORKERS=8
 CONTRA_ADMIN_KEYS=admin_pubkey
 
 # Gateway

.env.local

@@ -20,7 +20,11 @@ CONTRA_SIGVERIFY_QUEUE_SIZE=10000000
 CONTRA_SIGVERIFY_WORKERS=32
 CONTRA_WRITE_MAX_CONNECTIONS=1000000
 CONTRA_READ_MAX_CONNECTIONS=100000
+CONTRA_PG_MAX_CONNECTIONS=32
 CONTRA_MAX_TX_PER_BATCH=64
+CONTRA_BATCH_DEADLINE_MS=10
+CONTRA_BATCH_CHANNEL_CAPACITY=16
+CONTRA_MAX_SVM_WORKERS=8
 CONTRA_ADMIN_KEYS=your_admin_public_key # Public key of keypairs/admin.json
 
 # Gateway
@@ -32,6 +36,10 @@ GATEWAY_URL=http://gateway:8899
 # Local validator RPC
 LOCAL_VALIDATOR_RPC_URL=http://validator:8899
 
+# Auth service
+AUTH_PORT=8903
+JWT_SECRET=
+
 # Grafana
 GF_ADMIN_PASSWORD=admin
 

Cargo.toml

@@ -26,6 +26,7 @@ package.edition = "2021"
 [workspace.dependencies]
 anyhow = "1.0.100"
 argon2 = "0.5"
+async-channel = "2"
 base64 = "0.21"
 bincode = "1.3.3"
 borsh = "1.5.5"
@@ -117,7 +118,6 @@ testcontainers = "0.25"
 testcontainers-modules = { version = "0.13", features = ["postgres", "redis"] }
 thiserror = "2.0.17"
 tokio = { version = "=1.47.1", features = ["full"] }
-tokio-mpmc = "0.2.4"
 tokio-util = { version = "0.7", features = ["full"] }
 tracing = "0.1.41"
 tracing-subscriber = { version = "0.3", features = ["env-filter", "json"] }

bench-tps/.env.sample

@@ -20,8 +20,14 @@ POSTGRES_REPLICATION_PASSWORD=repl_password
 CONTRA_WRITE_PORT=8899
 CONTRA_SIGVERIFY_QUEUE_SIZE=1000
 CONTRA_SIGVERIFY_WORKERS=4
-CONTRA_WRITE_MAX_CONNECTIONS=100
+CONTRA_WRITE_MAX_CONNECTIONS=2048
+# Postgres pool size per node (default 32, capped at 256).
+CONTRA_PG_MAX_CONNECTIONS=32
 CONTRA_MAX_TX_PER_BATCH=256
+# Sequencer batch-flush deadline (ms).
+CONTRA_BATCH_DEADLINE_MS=10
+# Executor parallel SVM worker cap.
+CONTRA_MAX_SVM_WORKERS=8
 
 # Space-separated list of base58 admin public keys the node will accept.
 # At least one key is required for the bench setup phase (mint init, ATA, mint-to).
@@ -90,10 +96,14 @@ BENCH_ADMIN_KEYPAIR=
 BENCH_DURATION=60
 
 # Concurrent sender threads per flow
-BENCH_THREADS=3
+BENCH_THREADS=16
+
+# Transactions per batch produced by the generator; each batch is sent
+# concurrently by one sender task via join_all. Tune independently of BENCH_THREADS.
+BENCH_BATCH_SIZE=200
 
 # Milliseconds each sender thread sleeps between batches; 0 = maximum throughput
-BENCH_SENDER_SLEEP_MS=5
+BENCH_SENDER_SLEEP_MS=0
 
 # Raw token units minted to each account's ATA during setup; 1 unit consumed per tx
 BENCH_INITIAL_BALANCE=1000000

bench-tps/.env.sample.devnet

@@ -20,8 +20,14 @@ POSTGRES_REPLICATION_PASSWORD=repl_password
 CONTRA_WRITE_PORT=8899
 CONTRA_SIGVERIFY_QUEUE_SIZE=1000
 CONTRA_SIGVERIFY_WORKERS=4
-CONTRA_WRITE_MAX_CONNECTIONS=100
+CONTRA_WRITE_MAX_CONNECTIONS=2048
+# Postgres pool size per node (default 32, capped at 256).
+CONTRA_PG_MAX_CONNECTIONS=32
 CONTRA_MAX_TX_PER_BATCH=256
+# Sequencer batch-flush deadline (ms).
+CONTRA_BATCH_DEADLINE_MS=10
+# Executor parallel SVM worker cap.
+CONTRA_MAX_SVM_WORKERS=8
 
 # Space-separated list of base58 admin public keys the node will accept.
 # At least one key is required for the bench setup phase (mint init, ATA, mint-to).
@@ -106,10 +112,14 @@ BENCH_ADMIN_KEYPAIR=
 BENCH_DURATION=60
 
 # Concurrent sender threads per flow
-BENCH_THREADS=3
+BENCH_THREADS=16
+
+# Transactions per batch produced by the generator; each batch is sent
+# concurrently by one sender task via join_all. Tune independently of BENCH_THREADS.
+BENCH_BATCH_SIZE=200
 
 # Milliseconds each sender thread sleeps between batches; 0 = maximum throughput
-BENCH_SENDER_SLEEP_MS=5
+BENCH_SENDER_SLEEP_MS=0
 
 # Raw token units minted to each account's ATA during setup; 1 unit consumed per tx
 BENCH_INITIAL_BALANCE=1000000

bench-tps/Cargo.toml

@@ -9,6 +9,7 @@ path = "src/main.rs"
 
 [dependencies]
 anyhow = "1.0"
+async-channel = { workspace = true }
 clap = { workspace = true }
 contra-core = { workspace = true }
 contra-metrics = { workspace = true }

bench-tps/scripts/run.sh

@@ -57,6 +57,12 @@ BENCH_ENV="${BENCH_DIR}/.env"
 #                       script refreshes them to reload scrape config and
 #                       dashboards (safe; does not delete data volumes).
 #
+# --no-teardown         Skip `docker compose down` at exit.  Containers keep
+#                       running after the bench finishes (or is interrupted)
+#                       so logs and metrics can be collected for offline
+#                       analysis.  Stop them manually afterwards with
+#                       `docker compose -f <repo>/docker-compose.yml down`.
+#
 # --contra-threads N  Pin Contra service containers to the first N CPU cores
 #                     and the bench binary to the remaining cores.  When
 #                     omitted the default 75% / 25% split is used.
@@ -67,6 +73,7 @@ BENCH_ENV="${BENCH_DIR}/.env"
 REBUILD=0
 CLEAN=1          # default: always wipe volumes because validator resets each run
 REFRESH_METRICS=1
+TEARDOWN=1       # default: tear down containers on exit
 CONTRA_THREADS=""   # explicit core count for services (optional)
 BENCH_ARGS=()
 SKIP_NEXT=0
@@ -80,6 +87,7 @@ for arg in "$@"; do
         --rebuild)        REBUILD=1 ;;
         --no-clean)       CLEAN=0 ;;
         --no-refresh-metrics) REFRESH_METRICS=0 ;;
+        --no-teardown)    TEARDOWN=0 ;;
         --contra-threads) SKIP_NEXT=1 ;;  # value is the next token
         *)                BENCH_ARGS+=("${arg}") ;;
     esac
@@ -673,6 +681,12 @@ cleanup() {
     [ "${_CLEANUP_DONE}" -eq 1 ] && return
     _CLEANUP_DONE=1
     echo ""
+    if [ "${TEARDOWN}" -eq 0 ]; then
+        echo "Skipping teardown (--no-teardown). Containers are still running."
+        echo "  Inspect logs:  docker logs <container>"
+        echo "  Stop later:    docker compose -f ${REPO_ROOT}/docker-compose.yml down"
+        return
+    fi
     echo "Tearing down all services..."
     "${COMPOSE[@]}" down 2>/dev/null || true
     echo "Done."

bench-tps/src/args.rs

@@ -71,21 +71,30 @@ pub struct TransferArgs {
     /// Each account gets its own keypair, ATA, and initial token balance.
     /// Must be >= `--threads` to avoid multiple senders sharing a keypair
     /// (which would cause nonce conflicts).
-    #[arg(long, default_value_t = 50, env = "BENCH_ACCOUNTS")]
+    #[arg(long, default_value_t = 200, env = "BENCH_ACCOUNTS")]
     pub accounts: usize,
 
     /// Duration of the load phase in seconds.
     #[arg(long, default_value_t = 60, env = "BENCH_DURATION")]
     pub duration: u64,
 
-    /// Number of concurrent sender threads.
+    /// Number of concurrent sender tasks.
     ///
-    /// Each sender thread runs a blocking loop: pop batch → send each tx →
-    /// sleep `--sender-sleep-ms` → repeat.  More threads = higher throughput
-    /// up to the point where the node or network becomes the bottleneck.
-    #[arg(long, default_value_t = 4, env = "BENCH_THREADS")]
+    /// Each sender task runs an async loop: pop batch → send all txs
+    /// concurrently via `join_all` → sleep `--sender-sleep-ms` → repeat.
+    /// More tasks = higher throughput up to the point where the node or
+    /// network becomes the bottleneck.
+    #[arg(long, default_value_t = 16, env = "BENCH_THREADS")]
     pub threads: usize,
 
+    /// Number of transactions per batch produced by the generator.
+    ///
+    /// Each batch is sent concurrently by a single sender task using
+    /// `join_all`, so larger batches increase per-task parallelism.
+    /// Decoupled from `--threads` to allow independent tuning.
+    #[arg(long, default_value_t = 200, env = "BENCH_BATCH_SIZE")]
+    pub batch_size: usize,
+
     /// Number of distinct receiver accounts.
     ///
     /// Accounts are split into a sender pool (first half) and a receiver pool
@@ -115,7 +124,7 @@ pub struct TransferArgs {
     ///
     /// Use this to throttle the send rate without reducing `--threads`.
     /// A value of 0 disables the sleep entirely (maximum throughput mode).
-    #[arg(long, default_value_t = 5, env = "BENCH_SENDER_SLEEP_MS")]
+    #[arg(long, default_value_t = 0, env = "BENCH_SENDER_SLEEP_MS")]
     pub sender_sleep_ms: u64,
 
     /// Tracing log level.  One of: error, warn, info, debug, trace.