|
| 1 | +#!/usr/bin/env npx tsx |
| 2 | +/** |
| 3 | + * Local micro-benchmark — measures core code paths without wrangler dev. |
| 4 | + * |
| 5 | + * Tests the specific optimizations: |
| 6 | + * 1. Null bitmap decode (fast path for 0xFF/0x00 bytes) |
| 7 | + * 2. Coalesce range merging (autoCoalesceGap + coalesceRanges) |
| 8 | + * 3. canSkipPage (page-level filter pushdown) |
| 9 | + * 4. Operator pipeline (ScanOperator → FilterOperator → TopK) |
| 10 | + * 5. In-memory query via MaterializedExecutor |
| 11 | + * |
| 12 | + * Usage: npx tsx scripts/bench-local.ts |
| 13 | + */ |
| 14 | + |
| 15 | +import { decodePage, canSkipPage } from "../src/decode.js"; |
| 16 | +import { coalesceRanges, autoCoalesceGap, type Range } from "../src/coalesce.js"; |
| 17 | +import { QueryMode } from "../src/local.js"; |
| 18 | + |
| 19 | +// --------------------------------------------------------------------------- |
| 20 | +// Helpers |
| 21 | +// --------------------------------------------------------------------------- |
| 22 | + |
| 23 | +function timeIt(name: string, fn: () => void, iterations = 1000): { name: string; totalMs: number; opsPerSec: number; avgUs: number } { |
| 24 | + // Warmup |
| 25 | + for (let i = 0; i < Math.min(50, iterations); i++) fn(); |
| 26 | + |
| 27 | + const start = performance.now(); |
| 28 | + for (let i = 0; i < iterations; i++) fn(); |
| 29 | + const totalMs = performance.now() - start; |
| 30 | + const avgUs = (totalMs / iterations) * 1000; |
| 31 | + const opsPerSec = Math.round(iterations / (totalMs / 1000)); |
| 32 | + return { name, totalMs, opsPerSec, avgUs }; |
| 33 | +} |
| 34 | + |
| 35 | +async function timeItAsync(name: string, fn: () => Promise<void>, iterations = 100): Promise<{ name: string; totalMs: number; opsPerSec: number; avgUs: number }> { |
| 36 | + // Warmup |
| 37 | + for (let i = 0; i < Math.min(10, iterations); i++) await fn(); |
| 38 | + |
| 39 | + const start = performance.now(); |
| 40 | + for (let i = 0; i < iterations; i++) await fn(); |
| 41 | + const totalMs = performance.now() - start; |
| 42 | + const avgUs = (totalMs / iterations) * 1000; |
| 43 | + const opsPerSec = Math.round(iterations / (totalMs / 1000)); |
| 44 | + return { name, totalMs, opsPerSec, avgUs }; |
| 45 | +} |
| 46 | + |
| 47 | +type BenchResult = { name: string; totalMs: number; opsPerSec: number; avgUs: number }; |
| 48 | + |
| 49 | +function printTable(results: BenchResult[]): void { |
| 50 | + console.log( |
| 51 | + "\n" + |
| 52 | + [ |
| 53 | + "Benchmark".padEnd(55), |
| 54 | + "avg".padStart(12), |
| 55 | + "ops/sec".padStart(12), |
| 56 | + "total".padStart(10), |
| 57 | + ].join(" | ") |
| 58 | + ); |
| 59 | + console.log("-".repeat(95)); |
| 60 | + for (const r of results) { |
| 61 | + const avgStr = r.avgUs < 1000 ? `${r.avgUs.toFixed(1)}µs` : `${(r.avgUs / 1000).toFixed(2)}ms`; |
| 62 | + console.log( |
| 63 | + [ |
| 64 | + r.name.padEnd(55), |
| 65 | + avgStr.padStart(12), |
| 66 | + r.opsPerSec.toLocaleString().padStart(12), |
| 67 | + `${r.totalMs.toFixed(0)}ms`.padStart(10), |
| 68 | + ].join(" | ") |
| 69 | + ); |
| 70 | + } |
| 71 | + console.log("-".repeat(95)); |
| 72 | +} |
| 73 | + |
| 74 | +// --------------------------------------------------------------------------- |
| 75 | +// 1. Null bitmap decode |
| 76 | +// --------------------------------------------------------------------------- |
| 77 | + |
| 78 | +function benchBitmapDecode(): BenchResult[] { |
| 79 | + const results: BenchResult[] = []; |
| 80 | + const rowCount = 100_000; |
| 81 | + |
| 82 | + // All valid (0xFF bytes) — fast path should fly |
| 83 | + const allValid = new ArrayBuffer(8 + Math.ceil(rowCount / 8) + rowCount * 8); |
| 84 | + new Uint8Array(allValid, 0, Math.ceil(rowCount / 8)).fill(0xFF); |
| 85 | + // Write int64 data after bitmap |
| 86 | + const dv = new DataView(allValid, Math.ceil(rowCount / 8)); |
| 87 | + for (let i = 0; i < rowCount; i++) dv.setBigInt64(i * 8, BigInt(i), true); |
| 88 | + results.push(timeIt("bitmap: 100K rows, all valid (0xFF fast path)", () => { |
| 89 | + decodePage(allValid, "int64", 0, rowCount); |
| 90 | + }, 200)); |
| 91 | + |
| 92 | + // 50% null (alternating 0xAA bytes) |
| 93 | + const halfNull = new ArrayBuffer(Math.ceil(rowCount / 8) + rowCount * 8); |
| 94 | + new Uint8Array(halfNull, 0, Math.ceil(rowCount / 8)).fill(0xAA); // 10101010 |
| 95 | + results.push(timeIt("bitmap: 100K rows, 50% null (bit-by-bit)", () => { |
| 96 | + decodePage(halfNull, "int64", rowCount / 2, rowCount); |
| 97 | + }, 200)); |
| 98 | + |
| 99 | + // All null (0x00 bytes) — fast path should batch-add |
| 100 | + const allNull = new ArrayBuffer(Math.ceil(rowCount / 8) + rowCount * 8); |
| 101 | + new Uint8Array(allNull, 0, Math.ceil(rowCount / 8)).fill(0x00); |
| 102 | + results.push(timeIt("bitmap: 100K rows, all null (0x00 fast path)", () => { |
| 103 | + decodePage(allNull, "int64", rowCount, rowCount); |
| 104 | + }, 200)); |
| 105 | + |
| 106 | + return results; |
| 107 | +} |
| 108 | + |
| 109 | +// --------------------------------------------------------------------------- |
| 110 | +// 2. Coalesce ranges |
| 111 | +// --------------------------------------------------------------------------- |
| 112 | + |
| 113 | +function benchCoalesce(): BenchResult[] { |
| 114 | + const results: BenchResult[] = []; |
| 115 | + |
| 116 | + // Dense ranges (small gaps — should merge aggressively) |
| 117 | + const denseRanges: Range[] = []; |
| 118 | + for (let i = 0; i < 500; i++) { |
| 119 | + denseRanges.push({ column: `col${i % 5}`, offset: i * 8200, length: 8000 }); |
| 120 | + } |
| 121 | + results.push(timeIt("coalesce: 500 dense ranges (200B gaps)", () => { |
| 122 | + const gap = autoCoalesceGap(denseRanges); |
| 123 | + coalesceRanges(denseRanges, gap); |
| 124 | + }, 5000)); |
| 125 | + |
| 126 | + // Sparse ranges (large gaps — should keep separate) |
| 127 | + const sparseRanges: Range[] = []; |
| 128 | + for (let i = 0; i < 500; i++) { |
| 129 | + sparseRanges.push({ column: `col${i % 5}`, offset: i * 1_000_000, length: 8000 }); |
| 130 | + } |
| 131 | + results.push(timeIt("coalesce: 500 sparse ranges (992KB gaps)", () => { |
| 132 | + const gap = autoCoalesceGap(sparseRanges); |
| 133 | + coalesceRanges(sparseRanges, gap); |
| 134 | + }, 5000)); |
| 135 | + |
| 136 | + // Mixed — realistic scenario |
| 137 | + const mixedRanges: Range[] = []; |
| 138 | + for (let i = 0; i < 200; i++) { |
| 139 | + // Clustered in groups of 10 |
| 140 | + const group = Math.floor(i / 10); |
| 141 | + const inGroup = i % 10; |
| 142 | + mixedRanges.push({ column: `col${i % 3}`, offset: group * 500_000 + inGroup * 10_000, length: 8000 }); |
| 143 | + } |
| 144 | + results.push(timeIt("coalesce: 200 mixed ranges (clustered)", () => { |
| 145 | + const gap = autoCoalesceGap(mixedRanges); |
| 146 | + coalesceRanges(mixedRanges, gap); |
| 147 | + }, 5000)); |
| 148 | + |
| 149 | + return results; |
| 150 | +} |
| 151 | + |
| 152 | +// --------------------------------------------------------------------------- |
| 153 | +// 3. canSkipPage (page-level filter pushdown) |
| 154 | +// --------------------------------------------------------------------------- |
| 155 | + |
| 156 | +function benchCanSkipPage(): BenchResult[] { |
| 157 | + const results: BenchResult[] = []; |
| 158 | + const pages = Array.from({ length: 100 }, (_, i) => ({ |
| 159 | + byteOffset: BigInt(i * 80000), |
| 160 | + byteLength: 80000, |
| 161 | + rowCount: 10000, |
| 162 | + minValue: i * 10000, |
| 163 | + maxValue: (i + 1) * 10000 - 1, |
| 164 | + })); |
| 165 | + |
| 166 | + const filters = [{ column: "id", op: "gt" as const, value: 500000 }]; |
| 167 | + |
| 168 | + results.push(timeIt("canSkipPage: 100 pages × gt filter (50% skip)", () => { |
| 169 | + let skipped = 0; |
| 170 | + for (const page of pages) { |
| 171 | + if (canSkipPage(page, filters, "id")) skipped++; |
| 172 | + } |
| 173 | + }, 10000)); |
| 174 | + |
| 175 | + const rangeFilters = [ |
| 176 | + { column: "id", op: "gte" as const, value: 200000 }, |
| 177 | + { column: "id", op: "lt" as const, value: 800000 }, |
| 178 | + ]; |
| 179 | + results.push(timeIt("canSkipPage: 100 pages × range filter (40% skip)", () => { |
| 180 | + let skipped = 0; |
| 181 | + for (const page of pages) { |
| 182 | + if (canSkipPage(page, rangeFilters, "id")) skipped++; |
| 183 | + } |
| 184 | + }, 10000)); |
| 185 | + |
| 186 | + return results; |
| 187 | +} |
| 188 | + |
| 189 | +// --------------------------------------------------------------------------- |
| 190 | +// 4. In-memory query pipeline (MaterializedExecutor via fromJSON) |
| 191 | +// --------------------------------------------------------------------------- |
| 192 | + |
| 193 | +async function benchInMemoryQuery(): Promise<BenchResult[]> { |
| 194 | + const results: BenchResult[] = []; |
| 195 | + |
| 196 | + // Generate 10K rows |
| 197 | + const data = Array.from({ length: 10_000 }, (_, i) => ({ |
| 198 | + id: i, |
| 199 | + value: Math.random() * 1000, |
| 200 | + category: ["alpha", "beta", "gamma", "delta", "epsilon"][i % 5], |
| 201 | + region: ["us", "eu", "asia"][i % 3], |
| 202 | + })); |
| 203 | + |
| 204 | + const qm = QueryMode.fromJSON(data, "bench_data"); |
| 205 | + |
| 206 | + // Full scan |
| 207 | + results.push(await timeItAsync("query: 10K full scan", async () => { |
| 208 | + await qm.select("id", "value", "category").collect(); |
| 209 | + }, 500)); |
| 210 | + |
| 211 | + // Filter + collect |
| 212 | + results.push(await timeItAsync("query: 10K filter id>5000 (50% sel)", async () => { |
| 213 | + await qm.filter("id", "gt", 5000).collect(); |
| 214 | + }, 500)); |
| 215 | + |
| 216 | + // Filter + sort + limit (TopK) |
| 217 | + results.push(await timeItAsync("query: 10K filter+sort+limit(10)", async () => { |
| 218 | + await qm.filter("id", "gt", 5000).sort("value", "desc").limit(10).collect(); |
| 219 | + }, 500)); |
| 220 | + |
| 221 | + // Projection only |
| 222 | + results.push(await timeItAsync("query: 10K project 1 of 4 cols", async () => { |
| 223 | + await qm.select("id").collect(); |
| 224 | + }, 500)); |
| 225 | + |
| 226 | + // 100K rows |
| 227 | + const bigData = Array.from({ length: 100_000 }, (_, i) => ({ |
| 228 | + id: i, |
| 229 | + amount: Math.random() * 10000, |
| 230 | + category: ["A", "B", "C", "D", "E"][i % 5], |
| 231 | + })); |
| 232 | + const bigQm = QueryMode.fromJSON(bigData, "big_bench"); |
| 233 | + |
| 234 | + results.push(await timeItAsync("query: 100K full scan", async () => { |
| 235 | + await bigQm.collect(); |
| 236 | + }, 50)); |
| 237 | + |
| 238 | + results.push(await timeItAsync("query: 100K filter+sort+limit(100)", async () => { |
| 239 | + await bigQm.filter("id", "gt", 50000).sort("amount", "desc").limit(100).collect(); |
| 240 | + }, 100)); |
| 241 | + |
| 242 | + results.push(await timeItAsync("query: 100K filter id>90000 (10% sel)", async () => { |
| 243 | + await bigQm.filter("id", "gt", 90000).collect(); |
| 244 | + }, 100)); |
| 245 | + |
| 246 | + return results; |
| 247 | +} |
| 248 | + |
| 249 | +// --------------------------------------------------------------------------- |
| 250 | +// 5. Int64 decode (pure decode path) |
| 251 | +// --------------------------------------------------------------------------- |
| 252 | + |
| 253 | +function benchInt64Decode(): BenchResult[] { |
| 254 | + const results: BenchResult[] = []; |
| 255 | + |
| 256 | + // 100K int64 values, no nulls |
| 257 | + const rowCount = 100_000; |
| 258 | + const buf = new ArrayBuffer(rowCount * 8); |
| 259 | + const dv = new DataView(buf); |
| 260 | + for (let i = 0; i < rowCount; i++) dv.setBigInt64(i * 8, BigInt(i), true); |
| 261 | + |
| 262 | + results.push(timeIt("decode: 100K int64 values (no nulls)", () => { |
| 263 | + decodePage(buf, "int64", 0, rowCount); |
| 264 | + }, 200)); |
| 265 | + |
| 266 | + // 100K float64 values |
| 267 | + const fbuf = new ArrayBuffer(rowCount * 8); |
| 268 | + const fdv = new DataView(fbuf); |
| 269 | + for (let i = 0; i < rowCount; i++) fdv.setFloat64(i * 8, i * 1.5, true); |
| 270 | + |
| 271 | + results.push(timeIt("decode: 100K float64 values (no nulls)", () => { |
| 272 | + decodePage(fbuf, "float64", 0, rowCount); |
| 273 | + }, 200)); |
| 274 | + |
| 275 | + return results; |
| 276 | +} |
| 277 | + |
| 278 | +// --------------------------------------------------------------------------- |
| 279 | +// Main |
| 280 | +// --------------------------------------------------------------------------- |
| 281 | + |
| 282 | +async function main(): Promise<void> { |
| 283 | + console.log("QueryMode Local Micro-Benchmark"); |
| 284 | + console.log(`Node ${process.version} | ${process.platform} ${process.arch}`); |
| 285 | + console.log("=".repeat(95)); |
| 286 | + |
| 287 | + const allResults: BenchResult[] = []; |
| 288 | + |
| 289 | + console.log("\n## Bitmap Decode"); |
| 290 | + allResults.push(...benchBitmapDecode()); |
| 291 | + |
| 292 | + console.log("\n## Int64/Float64 Decode"); |
| 293 | + allResults.push(...benchInt64Decode()); |
| 294 | + |
| 295 | + console.log("\n## Coalesce Ranges"); |
| 296 | + allResults.push(...benchCoalesce()); |
| 297 | + |
| 298 | + console.log("\n## Page Skip (canSkipPage)"); |
| 299 | + allResults.push(...benchCanSkipPage()); |
| 300 | + |
| 301 | + console.log("\n## In-Memory Query Pipeline"); |
| 302 | + allResults.push(...await benchInMemoryQuery()); |
| 303 | + |
| 304 | + console.log("\n\n" + "=".repeat(95)); |
| 305 | + console.log("FULL RESULTS"); |
| 306 | + printTable(allResults); |
| 307 | +} |
| 308 | + |
| 309 | +main().catch(console.error); |
0 commit comments