@aztec/p2p 0.0.1-commit.96bb3f7 → 0.0.1-commit.993d240

package/src/client/test/p2p_client.batch_tx_requester.bench.README.md

@@ -0,0 +1,197 @@
+# BatchTxRequester Benchmarks
+
+This benchmark suite measures **how quickly a proposer node can fetch missing transactions from P2P peers** when building a block proposal under several controlled "who-has-which-txs" distributions.
+
+## Purpose
+
+When proposing a block, the node may have a block proposal containing a list of `txHashes`, but may be **missing the full `Tx` objects** locally. The node must fetch those missing txs from peers via the P2P req/resp layer.
+
+This benchmark answers:
+
+- How long does it take to fetch **N missing txs** (N ∈ **{10, 50, 100, 500}**)?
+- How do different **peer availability patterns** affect performance?
+
+## Architecture
+
+The benchmark runs a small simulated network on localhost:
+
+```
+┌─────────────────────────────────────────────────────────────────────┐
+│                        Test Process (Driver)                        │
+│   p2p_client.batch_tx_requester.bench.test.ts                       │
+│   ┌─────────────────────────────────────────────────────────────┐   │
+│   │            WorkerClientManager                              │   │
+│   │         (src/testbench/worker_client_manager.ts)            │   │
+│   └─────────────────────────────────────────────────────────────┘   │
+│                              │ IPC                                  │
+│         ┌────────────────────┼────────────────────┐                 │
+│         ▼                    ▼                    ▼                 │
+│   ┌───────────┐        ┌───────────┐        ┌───────────┐           │
+│   │ Worker 0  │◄──────►│ Worker 1  │◄──────►│ Worker N-1│           │
+│   │(Aggregator│  P2P   │(Responder)│  P2P   │(Responder)│           │
+│   │  Node)    │        │           │        │           │           │
+│   │ TxPool:[] │        │ TxPool:   │        │ TxPool:   │           │
+│   │           │        │ [txs...]  │        │ [txs...]  │           │
+│   └───────────┘        └───────────┘        └───────────┘           │
+└─────────────────────────────────────────────────────────────────────┘
+```
+
+- **N worker processes** are spawned using Node's **`child_process.fork()`**
+- Each worker runs a "light" **`P2PClient.Full`** node (enough to exercise the real P2P req/resp machinery)
+- Workers communicate with the test process via **IPC messages** defined in `p2p_client_testbench_worker.ts`
+
+### Why multiple processes?
+
+Using separate OS processes makes the setup closer to real networking behavior (independent event loops, scheduling, IPC boundaries), and avoids many artifacts you'd get by running everything in one process.
+
+### Worker roles
+
+The network is intentionally asymmetric:
+
+- **Worker 0 is the aggregator/proposer node**
+  - Starts with an **empty tx pool** (`[]`)
+  - Is the only worker instructed to run `BatchTxRequester` for each `BENCH_REQRESP` command
+- **Workers 1..N-1 are responder peers**
+  - Locally generate and filter txs according to the distribution pattern
+  - Respond to req/resp queries made by Worker 0's `BatchTxRequester`
+
+This models a proposer that has only `txHashes` in a proposal and must fetch the full tx bodies from the network.
+
+## Transaction Distribution Patterns
+
+Each benchmark case generates `missingTxCount` mock txs and assigns them to peers using one of these patterns:
+
+### `uniform`
+
+**Every responder peer has every transaction.**
+
+- Simulates the best-case: high replication / high gossip success
+- Expectation: the requester should quickly succeed; differences mostly reflect requester overhead and batching strategy
+
+### `sparse`
+
+**Each transaction exists on only a small subset of peers.**
+
+Each responder is bucketed and holds txs whose index falls into its bucket or the "next" bucket (striped by tx index).
+
+- Simulates partial propagation, churn, or uneven mempool convergence
+- Expectation: the requester must query multiple peers and cope with "misses"
+
+### `pinned-only`
+
+**Only a single "pinned" peer has the transactions; all other peers have none.**
+
+- Simulates "I know exactly who has the txs" (or a topology where one peer is the source of truth)
+- Useful to test whether "pinned peer" fast-paths work as intended
+
+> **Guardrail:** the pinned peer index must be within `(0, numberOfPeers)` (Worker 0 cannot be pinned).
+
+## Test Parameters
+
+| Parameter | Value | Description |
+|-----------|-------|-------------|
+| `PEERS_PER_RUN` | 30 | Number of worker processes spawned |
+| `MISSING_TX_COUNTS` | 10, 50, 100, 500 | Number of missing transactions to fetch |
+| `TIMEOUT_MS` | 30,000 ms | Per-case timeout for the requester |
+| `TEST_TIMEOUT_MS` | 600,000 ms | Overall Jest timeout (10 minutes) |
+
+## Running
+
+From the p2p package:
+
+```bash
+cd yarn-project/p2p
+yarn test src/client/test/p2p_client.batch_tx_requester.bench.test.ts
+```
+
+Or from repo root:
+
+```bash
+yarn test p2p_client.batch_tx_requester.bench.test.ts
+```
+
+The benchmark is intentionally long due to spawning many processes and running multiple cases.
+
+## Output Formats
+
+### Default: Markdown table to stdout
+
+If no env vars are set, the suite prints a table:
+
+```
+| Distribution | Missing | Duration (ms) | Fetched | Success |
+|--------------|---------|---------------|---------|---------|
+| pinned-only  |      10 |           123 |      10 |   Yes   |
+| pinned-only  |      50 |           145 |      50 |   Yes   |
+```
+
+### JSON metrics (for CI/dashboards)
+
+```bash
+BENCH_OUTPUT=/path/results.json yarn test ...
+```
+
+Writes JSON metrics like:
+- `BatchTxRequester/<distribution>/missing_<N>/duration` (ms)
+- `BatchTxRequester/<distribution>/missing_<N>/fetched` (txs)
+
+### Markdown file output
+
+```bash
+BENCH_OUTPUT_MD=/path/results.md yarn test ...
+```
+
+Writes the pretty table + summary to disk.
+
+## Interpreting Results
+
+For each case the benchmark records:
+
+- `durationMs`: wall-clock time spent inside the requester call
+- `fetchedCount`: how many txs were returned by the requester
+- `success`: `fetchedCount === missingTxCount`
+
+**Guidelines:**
+
+- **Always check `Success` first.** A faster run that fetched fewer txs is not a win.
+- Compare runs **within the same distribution + missing count** only.
+- Expect `pinned-only` to highlight pinned-peer behavior (fast if pinned peer is used effectively; slow if the algorithm wastes time sampling other peers).
+- Expect `sparse` to be the most "network-like" stress case, since many peers won't have each requested tx.
+
+## Determinism / Noise Reduction
+
+Inside each worker, the benchmark intentionally reduces variability:
+
+- **Unlimited rate limits** are installed so the req/resp rate limiter doesn't dominate results
+- **Deterministic tx generation** ensures all workers see the same tx set without large IPC payloads
+
+This makes the benchmark better for tracking regressions, but it is **not** a perfect model of production networking conditions.
+
+## Limitations
+
+This benchmark does **not** measure:
+
+- Real internet latency, NAT traversal, or adversarial peers (everything runs on localhost)
+- End-to-end proposer behavior beyond tx fetching (block construction, proving, etc.)
+- Gossip-based mempool convergence—tx availability is injected directly into each worker's in-memory tx pool
+
+## Files
+
+| File | Purpose |
+|------|---------|
+| `p2p_client.batch_tx_requester.bench.test.ts` | Test suite (cases, distributions, output formatting) |
+| `src/testbench/worker_client_manager.ts` | Worker process manager (forking, IPC, orchestration) |
+| `src/testbench/p2p_client_testbench_worker.ts` | General testbench worker implementation |
+| `src/test-helpers/testbench-utils.ts` | Shared mocks and utilities (InMemoryTxPool, InMemoryAttestationPool, etc.) |
+
+## Implementation Notes
+
+- Workers run TypeScript via `ts-node/esm` unless a compiled JS worker exists at `dest/testbench/p2p_client_testbench_worker.js`
+- Request/response rate limits are overridden so the benchmark is not throttled
+- Workers generate txs locally from a shared seed to avoid sending large tx payloads over IPC
+
+## Practical Tips
+
+- Run on an otherwise idle machine; CPU scheduling noise matters when spawning 30 node processes
+- If you see intermittent failures, increase `TIMEOUT_MS` or reduce `PEERS_PER_RUN` for local iteration
+- Use `BENCH_OUTPUT` in CI to track performance regressions over time