Merge pull request #52 from bnema/test/daemon-paint-benchmark-ack

test(daemon): keep paint benchmark past ack gate

commit: a386121bcb419250f311bf06e71a4425adc8f08b
author: bnema · 3mo ago · 2026-07-09 15:48 UTC
committer: GitHub · 3mo ago · 2026-07-09 15:48 UTC
merge of: ad25a70 + c31f501 · diff against the first parent
signature: GPG UnverifiedB5690EEEBB952194 The key is unknown

3 files · +7 -3

M AGENTS.md +1 -1


@@ -12,7 +12,7 @@ make test
12	12		go test ./internal/usecase/daemon -race -run TestName
13	13		make lint
14	14		make mocks
15		-	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc -bench=. -benchmem
	15	+	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem
16	16		```
17	17
18	18		- Format with goimports.


@@ -12,7 +12,7 @@ make test
12		go test ./internal/usecase/daemon -race -run TestName	12		go test ./internal/usecase/daemon -race -run TestName
13		make lint	13		make lint
14		make mocks	14		make mocks
15	-	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc -bench=. -benchmem	15	+	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem
16		```	16		```
17			17
18		- Format with goimports.	18		- Format with goimports.

M docs/performance.md +3 -2


@@ -5,7 +5,7 @@ Enable daemon pprof explicitly with `VEV_PPROF_ADDR=127.0.0.1:6060 vev --daemon`
5	5		Run microbenchmarks with:
6	6
7	7		```sh
8		-	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc -bench=. -benchmem
	8	+	go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem
9	9		```
10	10
11	11		Demo flood workloads are scripted in `scripts/bench-workloads.sh` for `yes`, `seq`, and `cat` styles for raw/tmux runs. vev does not currently support `vev new <name> -- command`, so automated vev workload command overrides are intentionally not documented. Capture bytes/frame and syscalls/frame externally (for example with `strace -c` around the client/daemon) and allocations with `-benchmem`/pprof.

@@ -24,12 +24,13 @@ strace -ff -c -o tmux-cat.strace tmux new-session -d -s perf-cat 'timeout 30s sh
24	24
25	25		Record host CPU, OS/kernel, terminal emulator, vev commit, tmux version, terminal size, workload duration, raw command lines, syscall summaries, bytes written per rendered frame, and heap profiles or allocation counts before comparing.
26	26
27		-	Microbenchmarks executed locally for this review using `go test ./internal/adapters/ipc -run '^$' -bench=. -benchmem`:
	27	+	Microbenchmarks executed locally for this review using `go test ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem`:
28	28
29	29		\| benchmark \| methodology \|
30	30		\| --- \| --- \|
31	31		\| BenchmarkTransportSend \| writes encoded frames to an in-memory discard `net.Conn`; allocations reported here are Send-side frame assembly costs. \|
32	32		\| BenchmarkTransportRecvReuse \| reads a pre-encoded frame from a looping in-memory reader; no `Send` call, goroutine, channel, or frame production work runs inside the measured loop, so allocations reported here are Recv-side costs only. \|
	33	+	\| BenchmarkPaintCachedSinglePaneDamage \| writes a small damage update into one pane, paints through the daemon compose→diff→send path, and advances the output ACK each iteration so the measured loop does not hit the unacked-output bailout. \|
33	34
34	35		\| workload \| vev baseline \| tmux baseline \| notes \|
35	36		\| --- \| --- \| --- \| --- \|

 Run microbenchmarks with:
 ```sh
-go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc -bench=. -benchmem
+go test ./pkg/renderer ./pkg/vt ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem
 ```
 Demo flood workloads are scripted in `scripts/bench-workloads.sh` for `yes`, `seq`, and `cat` styles for raw/tmux runs. vev does not currently support `vev new <name> -- command`, so automated vev workload command overrides are intentionally not documented. Capture bytes/frame and syscalls/frame externally (for example with `strace -c` around the client/daemon) and allocations with `-benchmem`/pprof.
 Record host CPU, OS/kernel, terminal emulator, vev commit, tmux version, terminal size, workload duration, raw command lines, syscall summaries, bytes written per rendered frame, and heap profiles or allocation counts before comparing.
-Microbenchmarks executed locally for this review using `go test ./internal/adapters/ipc -run '^$' -bench=. -benchmem`:
+Microbenchmarks executed locally for this review using `go test ./internal/adapters/ipc ./internal/usecase/daemon -run '^$' -bench=. -benchmem`:
 | benchmark | methodology |
 | --- | --- |
 | BenchmarkTransportSend | writes encoded frames to an in-memory discard `net.Conn`; allocations reported here are Send-side frame assembly costs. |
 | BenchmarkTransportRecvReuse | reads a pre-encoded frame from a looping in-memory reader; no `Send` call, goroutine, channel, or frame production work runs inside the measured loop, so allocations reported here are Recv-side costs only. |
+| BenchmarkPaintCachedSinglePaneDamage | writes a small damage update into one pane, paints through the daemon compose→diff→send path, and advances the output ACK each iteration so the measured loop does not hit the unacked-output bailout. |
 | workload | vev baseline | tmux baseline | notes |
 | --- | --- | --- | --- |

M internal/usecase/daemon/render_test.go +3 -0


@@ -574,6 +574,9 @@ func BenchmarkPaintCachedSinglePaneDamage(b *testing.B) {
574	574		for i := 0; i < b.N; i++ {
575	575		right.screen.Write([]byte("x\b"))
576	576		d.paint(sess, ac, false)
	577	+	// Mirror the client ACK pump so the steady state measures compose→diff→send,
	578	+	// not the ack-gate bailout after maxUnackedOutputStates frames.
	579	+	ac.advanceOutputAck(ac.nextStateNum)
577	580		}
578	581		}
579	582


@@ -574,6 +574,9 @@ func BenchmarkPaintCachedSinglePaneDamage(b *testing.B) {
574	for i := 0; i < b.N; i++ {	574		for i := 0; i < b.N; i++ {
575	right.screen.Write([]byte("x\b"))	575		right.screen.Write([]byte("x\b"))
576	d.paint(sess, ac, false)	576		d.paint(sess, ac, false)
		577	+	// Mirror the client ACK pump so the steady state measures compose→diff→send,
		578	+	// not the ack-gate bailout after maxUnackedOutputStates frames.
		579	+	ac.advanceOutputAck(ac.nextStateNum)
577	}	580		}
578	}	581		}
579		582