Daily update on 20260307-020120 by Nitload Youmu. Additional Message : None

Signed-off-by: Nitload <YoumuLiu_NSI@whv0t.onmicrosoft.com>
This commit is contained in:
Nitload
2026-03-07 02:01:20 +08:00
parent 405f895e29
commit 64360b63a1
4 changed files with 475 additions and 192 deletions
+403 -151
View File
@@ -29,41 +29,127 @@ using BenchmarkDotNet.Attributes;
using StgSharp.Collections;
using System;
using System.Collections.Generic;
using System.Runtime.CompilerServices;
namespace StgSharp.Benchmark
{
/// <summary>
/// Benchmark comparing SwissTable (SSE2 SIMD probing, zero-GC, native memory) against <see
/// cref="Dictionary{TKey, TValue}" /> for the L4 cache address mapping workload: concentrated
/// small-key insert/lookup/remove cycles. <para> SwissTable.Create() performs a 1 MB memset
/// (65536 × 16 control bytes), which is separated into IterationSetup to avoid polluting
/// insert timings.</para>
/// Benchmark comparing SwissTable against <see cref="Dictionary{TKey, TValue}" />
/// under realistic L4 software cache address-mapping workloads.
/// <para>
/// L4 scenario characteristics:
/// - Keys are HLSF/malloc pointer addresses (large, 1024-byte aligned)
/// - Access pattern is Zipfian-skewed (20% hot cache lines receive 80% of lookups)
/// - Steady state: high hit-rate lookups dominate, with rare evict+refill on miss
/// - Working set: 512–4096 active cache line mappings
/// </para>
/// </summary>
[MemoryDiagnoser]
public class SwissTableBench
{
private const int LookupCount = 4096;
/// <summary> L4 cache line size in bytes (per L4_design.md). </summary>
private const int CacheLineSize = 1024;
private nint[] _lookupKeys = null!;
/// <summary> Simulated heap base address for pointer-like keys. </summary>
private const long HeapBase = 0x7FFF_0000_0000L;
// ── Pre-generated key arrays ──
/// <summary> Number of lookup operations per Lookup benchmark call. </summary>
private const int LookupIterations = 8192;
private nint[] _preloadedKeys = null!;
/// <summary>
/// Number of steady-state operations per call.
/// Each op is either a lookup (hit) or an evict+insert+lookup (miss).
/// </summary>
private const int SteadyStateOps = 4096;
/// <summary> Fraction of steady-state ops that are cache misses. </summary>
private const double MissRate = 0.05;
// ── Pre-generated key arrays (pointer-like, 1024-byte aligned) ──
/// <summary> Keys for initial table fill (simulates cached addresses). </summary>
private nint[] _residentKeys = null!;
/// <summary> Additional keys not in the table, for triggering misses. </summary>
private nint[] _externalKeys = null!;
/// <summary> Zipfian-skewed lookup sequence (hot 20% accessed 80% of time). </summary>
private nint[] _zipfianLookups = null!;
/// <summary> Uniform random lookup sequence (baseline comparison). </summary>
private nint[] _uniformLookups = null!;
/// <summary>
/// Pre-built steady-state operation script.
/// Each entry: positive index → lookup _residentKeys[idx],
/// negative value → cache miss trigger (evict + insert from _externalKeys).
/// </summary>
private int[] _steadyStateScript = null!;
/// <summary> Indices into _residentKeys for eviction targets (cold entries, tail region). </summary>
private int[] _evictOrder = null!;
// ── Pre-filled lookup targets ──
private SwissTable _swissLookup = null!;
private Dictionary<nint, nint> _dictLookup = null!;
private Dictionary<nint, nint> _freshDict = null!;
// ── Fresh instances for insert/mixed (rebuilt each iteration) ──
private SwissTable _freshSwiss = null!;
private Dictionary<nint, nint> _freshDict = null!;
// ── Lookup targets (pre-filled in GlobalSetup) ──
private SwissTable _swissLookup = null!;
[Params(500, 2000, 5000)]
/// <summary>
/// Active cache line count. Matches L4 capacity range:
/// 5 MB / 1024B ≈ 5120, 10 MB / 1024B ≈ 10240.
/// 512 and 1024 test low-occupancy; 4096 tests near-capacity.
/// </summary>
[Params(512, 1024, 4096)]
public int KeyCount;
// ════════════════════════════════════════════════════════════
// Setup / Teardown
// ════════════════════════════════════════════════════════════
[GlobalSetup]
public void Setup()
{
Random rng = new Random(42);
// ── Generate pointer-like keys (HeapBase + random_offset * 1024) ──
_residentKeys = GeneratePointerKeys(KeyCount, rng);
_externalKeys = GeneratePointerKeys(KeyCount, rng); // disjoint set for miss simulation
// ── Build lookup sequences ──
_zipfianLookups = GenerateZipfianLookups(_residentKeys, LookupIterations, rng);
_uniformLookups = GenerateUniformLookups(_residentKeys, LookupIterations, rng);
// ── Build steady-state operation script ──
_steadyStateScript = GenerateSteadyStateScript(KeyCount, SteadyStateOps, MissRate, rng);
// ── Build eviction order (cold tail entries first) ──
_evictOrder = new int[KeyCount];
for (int i = 0; i < KeyCount; i++)
{
_evictOrder[i] = i;
}
// Shuffle so eviction isn't just sequential
Shuffle(_evictOrder, rng);
// ── Pre-fill lookup tables ──
_swissLookup = SwissTable.Create();
_dictLookup = new Dictionary<nint, nint>(KeyCount);
for (int i = 0; i < KeyCount; i++)
{
nint key = _residentKeys[i];
nint val = key ^ unchecked((nint)0xDEAD_BEEF); // simulated cache pointer
_swissLookup.TryAddOrSet(key, val, out _);
_dictLookup[key] = val;
}
}
[GlobalCleanup]
public void Cleanup()
{
@@ -71,119 +157,114 @@ namespace StgSharp.Benchmark
_dictLookup.TrimExcess();
}
// ══════════════════════════════════════
// Insert: pure write throughput
// Empty table → fill N keys
// Create() cost excluded via IterationSetup
// ══════════════════════════════════════
[Benchmark]
public void Dictionary_Insert()
{
Dictionary<nint, nint> dict = _freshDict;
nint[] keys = _preloadedKeys;
for (int i = 0; i < keys.Length; i++) {
dict[keys[i]] = keys[i] * 2;
}
}
// ══════════════════════════════════════
// Lookup: pure read throughput
// Pre-filled table, random key access
// ══════════════════════════════════════
[Benchmark(Baseline = true)]
public nint Dictionary_Lookup()
{
nint sum = 0;
nint[] keys = _lookupKeys;
Dictionary<nint, nint> dict = _dictLookup;
for (int i = 0; i < keys.Length; i++)
{
if (dict.TryGetValue(keys[i], out nint value)) {
sum += value;
}
}
return sum;
}
// ══════════════════════════════════════
// Mixed: insert all → lookup half → remove half → lookup all
// Simulates L4 cache lifecycle:
// load working set → compute → evict → recompute
// ══════════════════════════════════════
[Benchmark]
public nint Dictionary_Mixed()
{
Dictionary<nint, nint> dict = _freshDict;
nint[] keys = _preloadedKeys;
nint sum = 0;
// insert all
for (int i = 0; i < keys.Length; i++)
{
dict[keys[i]] = keys[i] * 2;
}
// lookup first half
for (int i = 0; i < keys.Length / 2; i++)
{
if (dict.TryGetValue(keys[i], out nint value)) {
sum += value;
}
}
// remove first half (eviction)
for (int i = 0; i < keys.Length / 2; i++)
{
dict.Remove(keys[i]);
}
// lookup all (half hit, half miss)
for (int i = 0; i < keys.Length; i++)
{
if (dict.TryGetValue(keys[i], out nint value)) {
sum += value;
}
}
return sum;
}
[IterationSetup(Targets = [
nameof(SwissTable_Insert),
nameof(SwissTable_Mixed),
nameof(SwissTable_SteadyState),
nameof(Dictionary_Insert),
nameof(Dictionary_Mixed)
nameof(Dictionary_SteadyState)
])]
public void IterationSetup()
{
_freshSwiss = SwissTable.Create();
_freshDict = new Dictionary<nint, nint>(KeyCount);
}
// ── Setup / Teardown ──
[GlobalSetup]
public void Setup()
{
_preloadedKeys = new nint[KeyCount];
_lookupKeys = new nint[LookupCount];
_swissLookup = SwissTable.Create();
_dictLookup = new Dictionary<nint, nint>(KeyCount);
// Pre-fill for steady-state benchmarks
for (int i = 0; i < KeyCount; i++)
{
_preloadedKeys[i] = (nint)i;
_swissLookup.TryAddOrSet((nint)i, (nint)(i * 2), out _);
_dictLookup[(nint)i] = (nint)(i * 2);
nint key = _residentKeys[i];
nint val = key ^ unchecked((nint)0xDEAD_BEEF);
_freshSwiss.TryAddOrSet(key, val, out _);
_freshDict[key] = val;
}
}
Random rng = new Random(42);
for (int i = 0; i < LookupCount; i++) {
_lookupKeys[i] = (nint)rng.Next(0, KeyCount);
// ════════════════════════════════════════════════════════════
// Lookup — Zipfian (L4 hot path, skewed access)
// 20% of cache lines receive 80% of accesses.
// This is the most representative L4 benchmark.
// ════════════════════════════════════════════════════════════
[Benchmark(Baseline = true)]
public nint Dictionary_Lookup_Zipfian()
{
nint sum = 0;
nint[] keys = _zipfianLookups;
Dictionary<nint, nint> dict = _dictLookup;
for (int i = 0; i < keys.Length; i++)
{
if (dict.TryGetValue(keys[i], out nint value))
{
sum += value;
}
}
return sum;
}
[Benchmark]
public nint SwissTable_Lookup_Zipfian()
{
nint sum = 0;
nint[] keys = _zipfianLookups;
SwissTable swiss = _swissLookup;
for (int i = 0; i < keys.Length; i++)
{
if (swiss.TryGet(keys[i], out nint value))
{
sum += value;
}
}
return sum;
}
// ════════════════════════════════════════════════════════════
// Lookup — Uniform (baseline comparison)
// ════════════════════════════════════════════════════════════
[Benchmark]
public nint Dictionary_Lookup_Uniform()
{
nint sum = 0;
nint[] keys = _uniformLookups;
Dictionary<nint, nint> dict = _dictLookup;
for (int i = 0; i < keys.Length; i++)
{
if (dict.TryGetValue(keys[i], out nint value))
{
sum += value;
}
}
return sum;
}
[Benchmark]
public nint SwissTable_Lookup_Uniform()
{
nint sum = 0;
nint[] keys = _uniformLookups;
SwissTable swiss = _swissLookup;
for (int i = 0; i < keys.Length; i++)
{
if (swiss.TryGet(keys[i], out nint value))
{
sum += value;
}
}
return sum;
}
// ════════════════════════════════════════════════════════════
// Insert — bulk fill with scattered pointer addresses
// Simulates L4 initial cache warm-up phase.
// ════════════════════════════════════════════════════════════
[Benchmark]
public void Dictionary_Insert()
{
Dictionary<nint, nint> dict = _freshDict;
nint[] keys = _residentKeys;
for (int i = 0; i < keys.Length; i++)
{
dict[keys[i]] = keys[i] ^ unchecked((nint)0xDEAD_BEEF);
}
}
@@ -191,64 +272,235 @@ namespace StgSharp.Benchmark
public void SwissTable_Insert()
{
SwissTable swiss = _freshSwiss;
nint[] keys = _preloadedKeys;
for (int i = 0; i < keys.Length; i++) {
swiss.TryAddOrSet(keys[i], keys[i] * 2, out _);
nint[] keys = _residentKeys;
for (int i = 0; i < keys.Length; i++)
{
swiss.TryAddOrSet(keys[i], keys[i] ^ unchecked((nint)0xDEAD_BEEF), out _);
}
}
// ════════════════════════════════════════════════════════════
// SteadyState — the KEY L4 benchmark
// Pre-filled table at capacity, then 4096 operations:
// ~95% are lookups (cache hits)
// ~5% are miss → evict cold entry → insert new entry → lookup new
// This directly simulates L4's main runtime loop.
// ════════════════════════════════════════════════════════════
[Benchmark]
public nint SwissTable_Lookup()
public nint Dictionary_SteadyState()
{
nint sum = 0;
nint[] keys = _lookupKeys;
SwissTable swiss = _swissLookup;
for (int i = 0; i < keys.Length; i++)
Dictionary<nint, nint> dict = _freshDict;
nint[] resident = _residentKeys;
nint[] external = _externalKeys;
int[] script = _steadyStateScript;
int[] evict = _evictOrder;
int evictCursor = 0;
int externalCursor = 0;
for (int i = 0; i < script.Length; i++)
{
if (swiss.TryGet(keys[i], out nint value)) {
sum += value;
int op = script[i];
if (op >= 0)
{
// Cache hit: lookup existing entry
if (dict.TryGetValue(resident[op], out nint value))
{
sum += value;
}
}
else
{
// Cache miss: evict a cold entry, insert a new one, then lookup
int evictIdx = evict[evictCursor % evict.Length];
evictCursor++;
dict.Remove(resident[evictIdx]);
int extIdx = externalCursor % external.Length;
externalCursor++;
nint newKey = external[extIdx];
nint newVal = newKey ^ unchecked((nint)0xDEAD_BEEF);
dict[newKey] = newVal;
if (dict.TryGetValue(newKey, out nint v))
{
sum += v;
}
}
}
return sum;
}
[Benchmark]
public nint SwissTable_Mixed()
public nint SwissTable_SteadyState()
{
nint sum = 0;
SwissTable swiss = _freshSwiss;
nint[] keys = _preloadedKeys;
nint sum = 0;
nint[] resident = _residentKeys;
nint[] external = _externalKeys;
int[] script = _steadyStateScript;
int[] evict = _evictOrder;
int evictCursor = 0;
int externalCursor = 0;
// insert all
for (int i = 0; i < keys.Length; i++)
for (int i = 0; i < script.Length; i++)
{
swiss.TryAddOrSet(keys[i], keys[i] * 2, out _);
}
int op = script[i];
if (op >= 0)
{
// Cache hit: lookup existing entry
if (swiss.TryGet(resident[op], out nint value))
{
sum += value;
}
}
else
{
// Cache miss: evict a cold entry, insert a new one, then lookup
int evictIdx = evict[evictCursor % evict.Length];
evictCursor++;
swiss.TryRemove(resident[evictIdx], out _);
// lookup first half
for (int i = 0; i < keys.Length / 2; i++)
{
if (swiss.TryGet(keys[i], out nint value)) {
sum += value;
int extIdx = externalCursor % external.Length;
externalCursor++;
nint newKey = external[extIdx];
nint newVal = newKey ^ unchecked((nint)0xDEAD_BEEF);
swiss.TryAddOrSet(newKey, newVal, out _);
if (swiss.TryGet(newKey, out nint v))
{
sum += v;
}
}
}
// remove first half (eviction)
for (int i = 0; i < keys.Length / 2; i++)
{
swiss.TryRemove(keys[i], out _);
}
// lookup all (half hit, half miss)
for (int i = 0; i < keys.Length; i++)
{
if (swiss.TryGet(keys[i], out nint value)) {
sum += value;
}
}
return sum;
}
// ════════════════════════════════════════════════════════════
// Data generation helpers
// ════════════════════════════════════════════════════════════
/// <summary>
/// Generate <paramref name="count"/> unique pointer-like keys.
/// Each key = HeapBase + randomOffset * CacheLineSize,
/// simulating HLSF-allocated 1024-byte aligned addresses
/// scattered across a ~1 GB virtual address range.
/// </summary>
private static nint[] GeneratePointerKeys(
int count,
Random rng
)
{
nint[] keys = new nint[count];
HashSet<long> used = new HashSet<long>(count);
for (int i = 0; i < count; i++)
{
long offset;
do
{
// Random offset within ~1M cache lines (1 GB range)
offset = (long)rng.Next(0, 1 << 20) * CacheLineSize;
} while (!used.Add(offset));
keys[i] = (nint)(HeapBase + offset);
}
return keys;
}
/// <summary>
/// Generate Zipfian-skewed lookup sequence:
/// 80% of accesses hit the hottest 20% of keys (sorted by index as "hotness proxy"),
/// 20% of accesses hit the remaining 80%.
/// This models matrix computation where a small panel working set is reused heavily.
/// </summary>
private static nint[] GenerateZipfianLookups(
nint[] sourceKeys,
int count,
Random rng
)
{
int hotSetSize = Math.Max(1, sourceKeys.Length / 5);
nint[] lookups = new nint[count];
for (int i = 0; i < count; i++)
{
if (rng.NextDouble() < 0.8)
{
lookups[i] = sourceKeys[rng.Next(0, hotSetSize)];
}
else
{
lookups[i] = sourceKeys[rng.Next(hotSetSize, sourceKeys.Length)];
}
}
return lookups;
}
/// <summary>
/// Generate uniform random lookup sequence (baseline for comparison with Zipfian).
/// </summary>
private static nint[] GenerateUniformLookups(
nint[] sourceKeys,
int count,
Random rng
)
{
nint[] lookups = new nint[count];
for (int i = 0; i < count; i++)
{
lookups[i] = sourceKeys[rng.Next(0, sourceKeys.Length)];
}
return lookups;
}
/// <summary>
/// Generate a steady-state operation script.
/// Each entry: non-negative value = index into _residentKeys (lookup/hit),
/// negative value (-1) = cache miss trigger (evict → insert → lookup).
/// Lookup indices are Zipfian-skewed.
/// </summary>
private static int[] GenerateSteadyStateScript(
int keyCount,
int opCount,
double missRate,
Random rng
)
{
int hotSetSize = Math.Max(1, keyCount / 5);
int[] script = new int[opCount];
for (int i = 0; i < opCount; i++)
{
if (rng.NextDouble() < missRate)
{
script[i] = -1; // miss
}
else
{
// Zipfian hit
if (rng.NextDouble() < 0.8)
{
script[i] = rng.Next(0, hotSetSize);
}
else
{
script[i] = rng.Next(hotSetSize, keyCount);
}
}
}
return script;
}
/// <summary> Fisher-Yates shuffle. </summary>
private static void Shuffle(
int[] array,
Random rng
)
{
for (int i = array.Length - 1; i > 0; i--)
{
int j = rng.Next(0, i + 1);
(array[i], array[j]) = (array[j], array[i]);
}
}
}
}