perf: increase thread-local heap capacity to reduce global refills

Increase kMaxMiniheapsPerShuffleVector from 24 to 48 and
kMiniheapRefillGoalSize from 4KB to 16KB to amortize the cost of
global heap operations.

In workloads with heavy thread churn (like larson benchmark), threads
frequently need to re-attach miniheaps when their shuffle vectors are
exhausted. Each global refill requires lock acquisition and freelist
walks.

The key insight is that kMaxMiniheapsPerShuffleVector is the dominant
factor: more attached miniheaps means more allocation capacity before
needing another global refill. For small 16-byte objects with 256 per
miniheap, 48 miniheaps provides ~12K allocations before refill vs ~6K
with the previous 24 miniheaps.

Testing showed diminishing returns on goal size once it exceeds what
the miniheaps array can provide, so 16KB (vs 256KB) captures most of
the benefit with lower RSS overhead.

Results on larson benchmark: relative time improved from ~33.5s to ~15.7s
(2.1x faster), throughput from ~29M to ~64M ops/sec.

Author: bpowers

src/common.h

@@ -163,8 +163,13 @@ static constexpr size_t kMinArenaExpansion = 4096;  // 4096 pages (16 MB on 4KB
 
 // ensures we amortize the cost of going to the global heap enough
 static constexpr uint64_t kMinStringLen = 8;
-static constexpr size_t kMiniheapRefillGoalSize = 4 * 1024;
-static constexpr size_t kMaxMiniheapsPerShuffleVector = 24;
+// Increased from 4KB to 16KB to reduce frequency of global refills.
+// Each refill grabs more capacity, trading some RSS for fewer lock acquisitions.
+static constexpr size_t kMiniheapRefillGoalSize = 16 * 1024;
+// Increased from 24 to 48 - this is the key driver of performance improvement.
+// More attached miniheaps means more capacity before needing global refills.
+// For small objects (256/miniheap): 48 miniheaps = 12K allocations before refill.
+static constexpr size_t kMaxMiniheapsPerShuffleVector = 48;
 
 // shuffle vector features
 static constexpr int16_t kMaxShuffleVectorLength = 1024;  // increased to support 16KB pages with 16-byte objects

src/meshable_arena.cc

@@ -134,7 +134,8 @@ void MeshableArena<PageSize>::afterForkChild() {
         }
 #endif
 
-        void *remapPtr = mmap(remove, sz, HL_MMAP_PROTECTION_MASK, kMapShared | MAP_FIXED, newFd, keepOff << kPageShift);
+        void *remapPtr =
+            mmap(remove, sz, HL_MMAP_PROTECTION_MASK, kMapShared | MAP_FIXED, newFd, keepOff << kPageShift);
         hard_assert_msg(remapPtr != MAP_FAILED, "mesh remap failed: %d", errno);
 
         return false;

src/meshable_arena.h

@@ -386,8 +386,7 @@ class MeshableArena : public mesh::OneWayMmapHeap {
 // Implementation
 
 template <size_t PageSize>
-MeshableArena<PageSize>::MeshableArena()
-    : SuperHeap(), _fastPrng(internal::seed(), internal::seed()) {
+MeshableArena<PageSize>::MeshableArena() : SuperHeap(), _fastPrng(internal::seed(), internal::seed()) {
   d_assert(getArenaInstance<PageSize>() == nullptr);
   getArenaInstance<PageSize>() = this;
 

src/mini_heap.h

@@ -196,7 +196,6 @@ class MiniHeap {
     d_assert(!_nextMeshed.hasValue());
   }
 
-
   inline Span span() const {
     return _span;
   }

src/size_class_reciprocals.h

@@ -32,31 +32,31 @@ static constexpr size_t kClassSizesMax = 25;
 namespace float_recip {
 
 inline constexpr float kReciprocals[kClassSizesMax] = {
-    1.0f / 16.0f,    // class 0: 16 bytes
-    1.0f / 16.0f,    // class 1: 16 bytes
-    1.0f / 32.0f,    // class 2: 32 bytes
-    1.0f / 48.0f,    // class 3: 48 bytes
-    1.0f / 64.0f,    // class 4: 64 bytes
-    1.0f / 80.0f,    // class 5: 80 bytes
-    1.0f / 96.0f,    // class 6: 96 bytes
-    1.0f / 112.0f,   // class 7: 112 bytes
-    1.0f / 128.0f,   // class 8: 128 bytes
-    1.0f / 160.0f,   // class 9: 160 bytes
-    1.0f / 192.0f,   // class 10: 192 bytes
-    1.0f / 224.0f,   // class 11: 224 bytes
-    1.0f / 256.0f,   // class 12: 256 bytes
-    1.0f / 320.0f,   // class 13: 320 bytes
-    1.0f / 384.0f,   // class 14: 384 bytes
-    1.0f / 448.0f,   // class 15: 448 bytes
-    1.0f / 512.0f,   // class 16: 512 bytes
-    1.0f / 640.0f,   // class 17: 640 bytes
-    1.0f / 768.0f,   // class 18: 768 bytes
-    1.0f / 896.0f,   // class 19: 896 bytes
-    1.0f / 1024.0f,  // class 20: 1024 bytes
-    1.0f / 2048.0f,  // class 21: 2048 bytes
-    1.0f / 4096.0f,  // class 22: 4096 bytes
-    1.0f / 8192.0f,  // class 23: 8192 bytes
-    1.0f / 16384.0f, // class 24: 16384 bytes
+    1.0f / 16.0f,     // class 0: 16 bytes
+    1.0f / 16.0f,     // class 1: 16 bytes
+    1.0f / 32.0f,     // class 2: 32 bytes
+    1.0f / 48.0f,     // class 3: 48 bytes
+    1.0f / 64.0f,     // class 4: 64 bytes
+    1.0f / 80.0f,     // class 5: 80 bytes
+    1.0f / 96.0f,     // class 6: 96 bytes
+    1.0f / 112.0f,    // class 7: 112 bytes
+    1.0f / 128.0f,    // class 8: 128 bytes
+    1.0f / 160.0f,    // class 9: 160 bytes
+    1.0f / 192.0f,    // class 10: 192 bytes
+    1.0f / 224.0f,    // class 11: 224 bytes
+    1.0f / 256.0f,    // class 12: 256 bytes
+    1.0f / 320.0f,    // class 13: 320 bytes
+    1.0f / 384.0f,    // class 14: 384 bytes
+    1.0f / 448.0f,    // class 15: 448 bytes
+    1.0f / 512.0f,    // class 16: 512 bytes
+    1.0f / 640.0f,    // class 17: 640 bytes
+    1.0f / 768.0f,    // class 18: 768 bytes
+    1.0f / 896.0f,    // class 19: 896 bytes
+    1.0f / 1024.0f,   // class 20: 1024 bytes
+    1.0f / 2048.0f,   // class 21: 2048 bytes
+    1.0f / 4096.0f,   // class 22: 4096 bytes
+    1.0f / 8192.0f,   // class 23: 8192 bytes
+    1.0f / 16384.0f,  // class 24: 16384 bytes
 };
 
 // Compute object index from byte offset using float reciprocal