0.35.4

docs/PERFORMANCE.md

@@ -70,7 +70,16 @@ Choose mimalloc only when all of these apply:
 - throughput under that turnover matters more than memory;
 - the server has at least 30% RAM headroom above the configured HOT cache size.
 
-Otherwise keep the default. The system allocator returns memory from evicted cache entries to the operating system right away, so process memory stays close to the configured cache size.
+Otherwise keep the default. The system allocator returns memory from evicted cache entries to the operating system shortly after eviction, so process memory stays close to the configured cache size.
+
+**Since 0.35.4** the default build returns freed memory to the operating system at most four times per second (plus the regular cache maintenance pass) instead of after every eviction. Most of the system allocator's slowdown under turnover came from these repeated trims, so the gap to mimalloc is now much smaller. Measured on the same 2 vCPU machine, one core, 256 MiB HOT cache filled with 100 KiB objects and every request a new URL:
+
+| Build | RSS after load | Throughput under cache turnover |
+|---|---:|---:|
+| 0.35.3, system allocator | 595 MiB | 2,900 req/s |
+| 0.35.4, system allocator | 639 MiB (+7%) | 7,000 req/s (2.4×) |
+
+The mimalloc rows above were measured with 0.35.2 and a different object mix; re-measure on your own workload before choosing the optional build.
 
 ## Production tuning