bench(memory): settle-until-quiescent GC barrier between iterations

benchmarks/memory/server/bench-utils.ts

@@ -17,14 +17,16 @@ export type RunSequentialRequestLoopOptions =
     iterations?: number
     buildRequest: (random: () => number, index: number) => Request
     validateResponse?: (response: Response, request: Request) => void
-    // For peak-shape scenarios only. Their signal is the footprint of a
-    // single request, but whether V8 collects iteration i's garbage before
-    // iteration i+1 allocates its payload is not reproducible run to run —
-    // the measured peak flips by a whole payload depending on GC timing.
-    // Forcing a collection between iterations pins the GC points so peak
-    // deterministically measures max(single-request footprint). Churn
-    // scenarios must never set this: their signal is accumulation across
-    // iterations, which a forced GC would mask.
+    // Whether V8 collects iteration i's garbage before iteration i+1
+    // allocates its payload is not reproducible run to run — the measured
+    // peak flips by a whole payload depending on GC timing, which shifts
+    // with runner hardware. Forcing a collection between iterations pins
+    // the GC points so peak deterministically measures the largest
+    // single-iteration footprint plus any reachable accumulation.
+    // Accumulation signals stay visible: leaked or cached objects are
+    // still referenced, so a forced collection cannot reclaim them — it
+    // only removes floating garbage, whose collection timing is the
+    // dominant cross-run noise source.
     pinGcBetweenIterations?: boolean
   }
 
@@ -102,13 +104,39 @@ export async function runSequentialRequestLoop(
     await drainResponse(response)
 
     if (pinGcBetweenIterations) {
-      // Two 0ms timer hops first, so trailing renderer/stream teardown
-      // scheduled for the next turns runs before the collection — otherwise
-      // its garbage survives into the next iteration's measurements.
-      await new Promise<void>((resolve) => setTimeout(resolve, 0))
-      await new Promise<void>((resolve) => setTimeout(resolve, 0))
-      // Present under CodSpeed (--expose-gc); absent in plain smoke runs.
-      ;(globalThis as { gc?: () => void }).gc?.()
+      await settleAndPinGc()
     }
   }
 }
+
+// Renderer/stream teardown spans several event-loop turns (React and Vue
+// need more than the two turns the original barrier allowed for), and a
+// too-short settle window leaks a whole payload of garbage past the
+// collection point, flipping the measured peak bimodally between runs.
+// Every step here is a fixed count on purpose: an adaptive exit (collect
+// until the post-GC heap size stops moving) makes the collection points
+// land at data-dependent turns, which re-introduces exactly the run-to-run
+// variance this barrier exists to remove — heap-size readings never fully
+// stabilize under the instrumented CI environment, so the exit point (and
+// with it every subsequent GC point) shifted between identical runs.
+const settleTurnsBeforeGc = 16
+
+// Settle trailing renderer/stream teardown with a fixed number of 0ms timer
+// hops (one full event-loop turn each, microtasks flushing between hops),
+// then pin a collection point. The second collection one turn later picks
+// up whatever the first collection's finalizers released. gc() is present
+// under CodSpeed (--expose-gc); in plain smoke runs there is nothing to
+// pin, so only the settle hops run.
+export async function settleAndPinGc() {
+  for (let turn = 0; turn < settleTurnsBeforeGc; turn++) {
+    await new Promise<void>((resolve) => setTimeout(resolve, 0))
+  }
+
+  const gc = (globalThis as { gc?: () => void }).gc
+
+  if (gc) {
+    gc()
+    await new Promise<void>((resolve) => setTimeout(resolve, 0))
+    gc()
+  }
+}

benchmarks/memory/server/scenarios/aborted-requests/react/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server aborted-requests (react)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/aborted-requests/shared.ts

@@ -1,3 +1,4 @@
+import { settleAndPinGc } from '#memory-server/bench-utils'
 import type { StartRequestHandler } from '#memory-server/bench-utils'
 
 export type { StartRequestHandler }
@@ -12,7 +13,7 @@ type AbortedRequestMode = {
   cancelMode: AbortedRequestCancelMode
 }
 
-const abortedRequestIterations = 100
+const abortedRequestIterations = 40
 let abortedRequestCounter = 0
 const eagerMarker = 'data-bench="aborted-requests-eager"'
 const alphaFallbackMarker = 'data-bench="aborted-requests-alpha-fallback"'
@@ -183,16 +184,12 @@ async function cancelReader(
   await reader.cancel()
 }
 
-const cancellationDrainTicks = 8
-
-// Fixed-count settlement barrier: each 0ms timer hop runs one full event-loop
-// turn (microtasks flush between hops), so post-abort renderer teardown
-// finishes within a deterministic number of turns instead of bleeding a
-// load-dependent amount of work into later iterations of the measured run.
+// Post-abort settlement barrier. Settles renderer teardown across a fixed
+// number of event-loop turns, then (under CodSpeed) pins a collection point
+// after every aborted request, so the measured peak cannot drift with how
+// much floating garbage the previous iterations happened to leave behind.
 async function drainCancellation() {
-  for (let tick = 0; tick < cancellationDrainTicks; tick++) {
-    await new Promise<void>((resolve) => setTimeout(resolve, 0))
-  }
+  await settleAndPinGc()
 }
 
 async function assertAbortedRequestsSanity(

benchmarks/memory/server/scenarios/aborted-requests/solid/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server aborted-requests (solid)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/aborted-requests/vue/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server aborted-requests (vue)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/error-paths/react/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server error-paths (react)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/error-paths/shared.ts

@@ -9,7 +9,7 @@ export type { StartRequestHandler }
 
 type Framework = 'react' | 'solid' | 'vue'
 
-const errorPathsIterations = 50
+const errorPathsIterations = 40
 const redirectSeed = 0xdecafbad
 const notFoundSeed = 0xdecafb0d
 const errorSeed = 0xdecafbed
@@ -142,6 +142,7 @@ export function createWorkloadGroup(
       iterations: errorPathsIterations,
       buildRequest: buildRedirectRequest,
       validateResponse: validateRedirectResponse,
+      pinGcBetweenIterations: true,
     })
 
   const runNotFound = () =>
@@ -150,6 +151,7 @@ export function createWorkloadGroup(
       iterations: errorPathsIterations,
       buildRequest: buildNotFoundRequest,
       validateResponse: validateNotFoundResponse,
+      pinGcBetweenIterations: true,
     })
 
   const runError = () =>
@@ -158,6 +160,7 @@ export function createWorkloadGroup(
       iterations: errorPathsIterations,
       buildRequest: buildErrorRequest,
       validateResponse: validateErrorResponse,
+      pinGcBetweenIterations: true,
     })
 
   const runUnmatched = () =>
@@ -166,6 +169,7 @@ export function createWorkloadGroup(
       iterations: errorPathsIterations,
       buildRequest: buildUnmatchedRequest,
       validateResponse: validateNotFoundResponse,
+      pinGcBetweenIterations: true,
     })
 
   return {

benchmarks/memory/server/scenarios/error-paths/solid/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server error-paths (solid)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/error-paths/vue/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server error-paths (vue)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/peak-large-page/react/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server peak-large-page (react)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/peak-large-page/solid/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server peak-large-page (solid)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/peak-large-page/vue/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server peak-large-page (vue)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/request-churn/react/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server request-churn (react)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/request-churn/shared.ts

@@ -10,7 +10,7 @@ export type { StartRequestHandler }
 type Framework = 'react' | 'solid' | 'vue'
 
 const benchmarkSeed = 0xdecafbad
-const requestChurnIterations = 200
+const requestChurnIterations = 40
 const itemPageMarker = 'data-bench="request-churn-item"'
 // Module-level so CodSpeed warmups and measurement never replay URLs.
 const benchmarkRandom = createDeterministicRandom(benchmarkSeed)
@@ -68,6 +68,7 @@ export function createWorkloadGroup(
       iterations: requestChurnIterations,
       buildRequest: buildItemRequest,
       validateResponse: validateItemResponse,
+      pinGcBetweenIterations: true,
     })
 
   return {

benchmarks/memory/server/scenarios/request-churn/solid/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server request-churn (solid)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/request-churn/vue/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server request-churn (vue)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/serialization-payload/react/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server serialization-payload (react)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/serialization-payload/shared.ts

@@ -9,7 +9,7 @@ export type { StartRequestHandler }
 type Framework = 'react' | 'solid' | 'vue'
 
 const benchmarkSeed = 0x51eaa11
-const serializationPayloadIterations = 20
+const serializationPayloadIterations = 12
 const payloadPageMarker = 'data-bench="serialization-payload"'
 
 const requestInit = {

benchmarks/memory/server/scenarios/serialization-payload/solid/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server serialization-payload (solid)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/serialization-payload/vue/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server serialization-payload (vue)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/server-fn-churn/react/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server server-fn-churn (react)',
     watch: false,
     environment: 'node',

benchmarks/memory/server/scenarios/server-fn-churn/shared.ts

@@ -38,7 +38,7 @@ type SerovalNode =
 const benchmarkSeed = 0xdecafbad
 const payloadSeed = 0x51f0cafe
 const fixtureCount = 16
-const serverFnChurnIterations = 150
+const serverFnChurnIterations = 40
 const origin = 'http://localhost'
 const tssContentTypeFramed = 'application/x-tss-framed'
 const acceptHeader = `${tssContentTypeFramed}, application/x-ndjson, application/json`
@@ -206,6 +206,7 @@ export async function createWorkloadGroup(
     runSequentialRequestLoop(handler, {
       seed: benchmarkSeed,
       iterations: serverFnChurnIterations,
+      pinGcBetweenIterations: true,
       buildRequest: (_random, index) => {
         const fixtureIndex = Math.floor(index / 2) % fixtureCount
 

benchmarks/memory/server/scenarios/server-fn-churn/solid/vite.config.ts

@@ -22,6 +22,11 @@ export default defineConfig({
     minify: false,
   },
   test: {
+    // Keep lazily-compiled code alive for the whole run: the pinned
+    // collections age code fast enough for V8 to flush unused bytecode,
+    // and the mid-measurement recompile injects a multi-MB allocation
+    // burst at a run-dependent time.
+    execArgv: ['--no-flush-bytecode'],
     name: '@benchmarks/memory-server server-fn-churn (solid)',
     watch: false,
     environment: 'node',