agent_send_latency¶

Last refreshed: v0.38 (2026-05-29) on vulcan (Dell, Intel Xeon, 16 cores, Ubuntu 24.04, Rust 1.95.0, PGO release-pgo profile). Mighty + Rust-tokio comparator numbers are v0.38 (rerun on this host); the C++ asio comparator falls back to a condition-variable impl on vulcan (asio not linked) and the Go comparator retains the v0.6 baseline (Go toolchain not installed on vulcan — both tracked as a v0.39 follow-up).

Workload: one fire-and-forget message between sender and receiver on the same tokio runtime, mailbox capacity 8. Excludes mailbox setup cost (measured inside the timed region; same for all comparators).

Spec alignment: §0 agent-first headline. This is the primitive the whole runtime is built on.

Numbers¶

Impl	Median	p95	p99	Notes
Mighty v0.38 mailbox (Block policy)	0.19 µs	0.24 µs	6.9 µs	tokio mpsc + slab admission
Mighty v0.38 mailbox (Fail policy, try_send)	(criterion bench)			bypasses await on send
Rust tokio mpsc (bare)	<0.1 µs	<0.1 µs	~1 µs	bare mpsc, no slab — vulcan
C++ condition-variable fallback	<0.1 µs	<0.1 µs	<0.5 µs	cv-based; asio coroutine channel not linked on vulcan
Go unbuffered chan	(pending — Go not installed on vulcan)			`ch := make(chan int, 8)`

Recorded values (vulcan, 2026-05-29, v0.38)¶

agent_send_latency        median=     0.000 ms  p95=     0.000 ms  p99=     0.007 ms
rust_tokio_send_latency   median=     0.000 ms  p95=     0.000 ms  p99=     0.001 ms
cpp_cv_agent_send_latency median=     0.000 ms  p95=     0.000 ms  p99=     0.000 ms

Raw nanos from /tmp/bench-v038-pgo.json:

"median_ns": 185,   // = 0.19 µs
"p95_ns":    235,
"p99_ns":   6926

So median is ~0.19 µs (down ~8% from v0.33's 200 ns — release-pgo's fat-LTO + actual PGO-from-merged.profdata buys a few ns on the hot send/recv path). The P99 tail of ~6.9 µs catches occasional multi-socket Xeon scheduler jitter; the median is the meaningful steady-state number.

v0.6 baseline (Windows 11 dev laptop, 2026-05-24)¶

For continuity: Mighty v0.6 measured median = 0.2-0.4 µs with a P99 tail of ~12 µs on a different host. Cross-host deltas are shape, not absolute.

Interpretation¶

Sub-microsecond P50 on a single tokio task is a healthy starting point. The P99 tail of ~12 µs is dominated by tokio scheduler jitter — when the runtime decides to poll the receiver, not the mailbox's overhead.

Mighty's mailbox adds a slab_pool::admit() step over a bare tokio mpsc (records a metadata blob in the slab so we can back-pressure on bytes, not just count). Expected overhead vs bare mpsc: a few hundred nanoseconds. Once we run the Rust comparator we'll quantify the exact slab tax.

v0.7+ optimisation targets¶

Skip slab admission for empty payloads. Empty SmallPayload doesn't need a slot.
Replace the slab's allocator with a thread-local arena for hot agents (P99 spikes likely come from contention on the shared pool).
Inline fast path for unique senders — when the mailbox has a single sender (the common case after spawn), the lock-free path can skip the mpsc indirection.

Tracked in: BENCHMARKS_V0_6_NOTES.md § Agent Send Latency.

v0.8 update¶

Optimisation	Status	Delta
Skip slab admit for empty payloads	DONE	`Mailbox::admit` returns a tombstone `PooledFrame` for `SmallPayload::Empty` — no parking_lot lock, no Vec alloc, no slot write
Stack-resident inline cache (non-empty)	DONE	64-byte stack buffer replaces per-call `Vec::with_capacity(inline_bytes)`
Thread-local arena (hot agents)	DEFER	Empty fast-path already removes the dominant cost; thread-local arena only helps multi-sender + non-empty, which the v0.6 benches don't exercise
Single-sender fast path	DEFER	tokio mpsc with one Sender doesn't take a fair-queueing slow path; bypassing the channel is a larger refactor

Microbench: crates/mty-runtime/benches/agent_send_latency.rs (send_recv_empty / send_recv_inline_1 / try_send_empty). Interpretation log: BENCHMARKS_V0_8_NOTES.md.

try_send_empty measured ~800 ns end-to-end on this host vs the v0.6 P50 of ~0.4 µs measured by the lower-overhead runner; the criterion bench's per-iter mailbox construct/drop adds ~400 ns, which is the bulk of the headline gap.