Add JMH benchmarks comparing DDCache, SimpleUtf8Cache, GenerationalUtf8Cache

dd-trace-core/src/jmh/java/datadog/trace/common/writer/ddagent/Utf8Benchmark.java

@@ -6,6 +6,8 @@
 
 import datadog.communication.serialization.GenerationalUtf8Cache;
 import datadog.communication.serialization.SimpleUtf8Cache;
+import datadog.trace.api.cache.DDCache;
+import datadog.trace.api.cache.DDCaches;
 import java.nio.charset.StandardCharsets;
 import org.openjdk.jmh.annotations.Benchmark;
 import org.openjdk.jmh.annotations.BenchmarkMode;
@@ -23,6 +25,11 @@
  * gc" to check bytes / op. Since {@link String#getBytes(java.nio.charset.Charset)} is intrinsified
  * the caches typically perform worse throughput wise; the benefit of the caches is to reduce
  * allocation.
+ *
+ * <p>{@link DDCache} (general-purpose, no eviction quality control, no recalibration) is included
+ * as a baseline "naive cache" arm alongside the purpose-built {@link SimpleUtf8Cache} / {@link
+ * GenerationalUtf8Cache}: it has no seen-once filtering, so every miss -- including one-time
+ * "custom" values -- immediately allocates and stores an entry, unlike the other two.
  */
 @BenchmarkMode(Mode.Throughput)
 public class Utf8Benchmark {
@@ -49,6 +56,14 @@ public static final byte[] tagUtf8_w_cache() {
     return tag.getBytes(StandardCharsets.UTF_8);
   }
 
+  static final DDCache<String, byte[]> DDCACHE_TAG_CACHE = DDCaches.newFixedSizeCache(128);
+
+  @Benchmark
+  public static final byte[] tagUtf8_w_ddcache() {
+    String tag = nextTag();
+    return DDCACHE_TAG_CACHE.computeIfAbsent(tag, t -> t.getBytes(StandardCharsets.UTF_8));
+  }
+
   @Benchmark
   public static final void valueUtf8_baseline(Blackhole bh) {
     for (int i = 0; i < NUM_LOOKUPS; ++i) {
@@ -92,6 +107,23 @@ public static final void valueUtf8_cache_simple(Blackhole bh) {
     }
   }
 
+  // Matches GENERATIONAL_VALUE_CACHE's combined eden+tenured capacity (64 + 128).
+  static final DDCache<String, byte[]> DDCACHE_VALUE_CACHE = DDCaches.newFixedSizeCache(192);
+
+  @Benchmark
+  public static final void valueUtf8_cache_ddcache(Blackhole bh) {
+    DDCache<String, byte[]> valueCache = DDCACHE_VALUE_CACHE;
+    // No recalibrate() call: DDCache has no decay/eviction-quality bookkeeping to drive.
+
+    for (int i = 0; i < NUM_LOOKUPS; ++i) {
+      String tag = nextTag();
+      String value = nextValue(tag);
+
+      byte[] lookup = valueCache.computeIfAbsent(value, v -> v.getBytes(StandardCharsets.UTF_8));
+      bh.consume(lookup);
+    }
+  }
+
   @Benchmark
   public static final void valueUtf8_nocache(Blackhole bh) {
     for (int i = 0; i < NUM_LOOKUPS; ++i) {

dd-trace-core/src/jmh/java/datadog/trace/common/writer/ddagent/Utf8CacheSizeBenchmark.java

@@ -0,0 +1,153 @@
+package datadog.trace.common.writer.ddagent;
+
+import static datadog.trace.common.writer.ddagent.Utf8Workload.NUM_LOOKUPS;
+import static datadog.trace.common.writer.ddagent.Utf8Workload.nextTag;
+import static datadog.trace.common.writer.ddagent.Utf8Workload.nextValue;
+import static java.nio.charset.StandardCharsets.UTF_8;
+
+import datadog.communication.serialization.EncodingCache;
+import datadog.communication.serialization.GenerationalUtf8Cache;
+import datadog.communication.serialization.SimpleUtf8Cache;
+import datadog.trace.api.cache.DDCache;
+import datadog.trace.api.cache.DDCaches;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.Param;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.infra.Blackhole;
+
+/**
+ * Sweeps cache capacity for all three caches against the same {@link Utf8Workload}, since the fixed
+ * sizes in {@link Utf8Benchmark} (128, 192) only show each cache at one operating point. The
+ * capacities picked -- 10, 100, 1000 -- roughly bracket the range actually used in production (a
+ * handful of near-static tag names up to {@link SimpleUtf8Cache#MAX_CAPACITY}), so this shows how
+ * each design degrades as the workload's ~180-entry stable set plus unbounded "custom" churn
+ * outgrows or comfortably fits the budget.
+ *
+ * <p>Best paired with {@code -prof gc}: the interesting signal is allocation (entries created,
+ * bytes/op), not throughput -- see {@link Utf8Benchmark}'s note on why raw ops/s undersells the
+ * caches relative to the intrinsified {@code getBytes} baseline.
+ *
+ * <p>{@code NOCACHE} is the uncached baseline (always calls {@code getBytes} directly) -- it
+ * ignores {@code capacity}, so it's redundant across the three capacity values, but keeping it in
+ * the same parameter sweep makes it trivial to eyeball against every row of the cached variants.
+ *
+ * <h2>Results (capacity sweep, {@code -prof gc}, normalized)</h2>
+ *
+ * <p>Throughput in ops/s, allocation in bytes/op, GC rate normalized to counts/sec ({@code
+ * gc.count}/{@code gc.time} are cumulative sums over the whole run, so raw values aren't comparable
+ * across runs with different fork counts without dividing by total measurement time).
+ *
+ * <pre>
+ * capacity  kind          ops/s     alloc B/op  alloc vs. NOCACHE  alloc MB/s  GC/s   GC time %
+ * 10        NOCACHE       2007.9    1,247,228   --                 2388.3     3.98   0.22%
+ * 10        DDCACHE       1315.7    1,447,111   +16.0% (worse)      1815.7    3.03   0.17%
+ * 10        SIMPLE        1401.6    1,305,537   +4.7%  (worse)      1745.1    2.91   0.16%
+ * 10        GENERATIONAL  1205.6    1,318,938   +5.7%  (worse)      1516.4    2.53   0.15%
+ * 100       NOCACHE       2003.6    1,247,238   --                 2383.1     3.96   0.22%
+ * 100       DDCACHE       1397.4    1,248,407   +0.1%  (breakeven)  1663.7    2.77   0.16%
+ * 100       SIMPLE        1451.3    1,161,541   -6.9%               1607.6    2.68   0.15%
+ * 100       GENERATIONAL  1422.4    1,046,133   -16.1%              1419.0    2.36   0.13%
+ * 1000      NOCACHE       1966.6    1,247,231   --                 2339.1     3.88   0.22%
+ * 1000      DDCACHE       1648.9    1,034,356   -17.1%              1626.6    2.72   0.20%
+ * 1000      SIMPLE        1641.3    974,316     -21.9%              1525.1    2.54   0.14%
+ * 1000      GENERATIONAL  1709.5    961,062     -23.0%              1566.8    2.61   0.15%
+ * </pre>
+ *
+ * <h2>Key findings</h2>
+ *
+ * <ul>
+ *   <li>All three caches lose on raw throughput to the uncached baseline at every capacity tested.
+ *       {@link java.lang.String#getBytes(java.nio.charset.Charset)} is intrinsified, so a cache
+ *       lookup is strictly more expensive per-op than just re-encoding -- these caches only pay off
+ *       on allocation, not throughput. Any cache meant to reduce app-thread allocation has to beat
+ *       a TLAB bump-pointer allocation, which is a very cheap operation to beat.
+ *   <li>Allocation reduction is real but smaller than naively expected -- at capacity 1000 all
+ *       three caches reduce allocation by only ~17-23%, not the ~90% you'd guess from hit rate
+ *       alone. Root cause: {@link Utf8Workload#nextStandardValue}/{@link
+ *       Utf8Workload#nextCustomValue} build each value via {@code tag + int} string concatenation
+ *       on <em>every</em> lookup, which allocates a fresh {@code String} regardless of whether the
+ *       subsequent UTF-8 encoding is cached. For ASCII-compact strings, {@code getBytes(UTF_8)} is
+ *       itself just a copy of the existing byte array, so a cache hit only saves that one array
+ *       copy -- not the string-construction cost. This structurally caps the max possible
+ *       allocation win for any UTF8 cache on this workload.
+ *   <li>{@code DDCACHE} has a consistent allocation disadvantage vs. {@code SIMPLE}/{@code
+ *       GENERATIONAL} (worst at capacity 10, where it's the only arm to allocate <em>more</em> than
+ *       {@code NOCACHE}), from its generic {@code Pair.of(key, value)} wrapper allocated on every
+ *       stored miss ({@code FixedSizeCache.produceAndStoreValue}). {@code SIMPLE}/{@code
+ *       GENERATIONAL} avoid this on one-off values via a marker/bloom-style gate ({@code
+ *       Caching.mark}) that defers the wrapper allocation to a value's <em>second</em> observed
+ *       touch -- same immutable-wrapper-for-lock-freedom pattern as {@code DDCACHE}, but the
+ *       wrapper cost is paid only for values that actually recur. The marker is advisory, not
+ *       authoritative: a wrong mark only costs a redundant allocation on the next touch, it can
+ *       never cause the cache to return incorrect data.
+ *   <li>{@code SIMPLE} and {@code GENERATIONAL} are close to equal at capacity 100, with {@code
+ *       GENERATIONAL} pulling ahead at capacity 1000 -- worth checking what capacity these caches
+ *       actually run at in production before assuming {@code SIMPLE}'s extra complexity over {@code
+ *       GENERATIONAL} is still earning its keep.
+ * </ul>
+ */
+@BenchmarkMode(Mode.Throughput)
+@State(Scope.Thread)
+public class Utf8CacheSizeBenchmark {
+
+  public enum CacheKind {
+    NOCACHE,
+    DDCACHE,
+    SIMPLE,
+    GENERATIONAL
+  }
+
+  @Param({"10", "100", "1000"})
+  int capacity;
+
+  @Param({"NOCACHE", "DDCACHE", "SIMPLE", "GENERATIONAL"})
+  CacheKind cacheKind;
+
+  EncodingCache cache;
+
+  // Only Simple/Generational have anything to recalibrate; DDCache is a no-op here.
+  Runnable recalibrate;
+
+  @Setup(Level.Trial)
+  public void setup() {
+    switch (cacheKind) {
+      case NOCACHE:
+        cache = value -> value.toString().getBytes(UTF_8);
+        recalibrate = () -> {};
+        break;
+      case DDCACHE:
+        DDCache<String, byte[]> ddCache = DDCaches.newFixedSizeCache(capacity);
+        cache = value -> ddCache.computeIfAbsent(value.toString(), v -> v.getBytes(UTF_8));
+        recalibrate = () -> {};
+        break;
+      case SIMPLE:
+        SimpleUtf8Cache simpleCache = new SimpleUtf8Cache(capacity);
+        cache = simpleCache;
+        recalibrate = simpleCache::recalibrate;
+        break;
+      case GENERATIONAL:
+        GenerationalUtf8Cache generationalCache = new GenerationalUtf8Cache(capacity);
+        cache = generationalCache;
+        recalibrate = generationalCache::recalibrate;
+        break;
+    }
+  }
+
+  @Benchmark
+  public void lookup(Blackhole bh) {
+    // Single thread drives recalibrate inline, at a transaction boundary -- see Utf8Benchmark.
+    recalibrate.run();
+
+    for (int i = 0; i < NUM_LOOKUPS; ++i) {
+      String tag = nextTag();
+      String value = nextValue(tag);
+
+      bh.consume(cache.encode(value));
+    }
+  }
+}

dd-trace-core/src/jmh/java/datadog/trace/common/writer/ddagent/Utf8ConcurrentBenchmark.java

@@ -6,6 +6,9 @@
 
 import datadog.communication.serialization.GenerationalUtf8Cache;
 import datadog.communication.serialization.SimpleUtf8Cache;
+import datadog.trace.api.cache.DDCache;
+import datadog.trace.api.cache.DDCaches;
+import java.nio.charset.StandardCharsets;
 import org.openjdk.jmh.annotations.Benchmark;
 import org.openjdk.jmh.annotations.BenchmarkMode;
 import org.openjdk.jmh.annotations.Group;
@@ -29,12 +32,17 @@
  *
  * <p>The recalibrate thread runs continuously, which is deliberately more aggressive than a real
  * periodic cadence — it maximizes contention so this doubles as a concurrency guardrail.
+ *
+ * <p>{@link DDCache} has no recalibrate step, so its group uses 8 lookup threads (no dedicated
+ * recalibrate thread) against the same shared cache -- a pure contention/throughput comparison
+ * rather than a recalibrate-race guardrail.
  */
 @BenchmarkMode(Mode.Throughput)
 @State(Scope.Group)
 public class Utf8ConcurrentBenchmark {
   static final GenerationalUtf8Cache VALUE_CACHE = new GenerationalUtf8Cache(64, 128);
   static final SimpleUtf8Cache SIMPLE_VALUE_CACHE = new SimpleUtf8Cache(128);
+  static final DDCache<String, byte[]> DDCACHE_VALUE_CACHE = DDCaches.newFixedSizeCache(192);
 
   @Benchmark
   @Group("generational")
@@ -69,4 +77,16 @@ public void simple_lookup(Blackhole bh) {
   public void simple_recalibrate() {
     SIMPLE_VALUE_CACHE.recalibrate();
   }
+
+  @Benchmark
+  @Group("ddcache")
+  @GroupThreads(8)
+  public void ddcache_lookup(Blackhole bh) {
+    for (int i = 0; i < NUM_LOOKUPS; ++i) {
+      String tag = nextTag();
+      String value = nextValue(tag);
+      bh.consume(
+          DDCACHE_VALUE_CACHE.computeIfAbsent(value, v -> v.getBytes(StandardCharsets.UTF_8)));
+    }
+  }
 }