Added benchmarks for `upb_Arena_Fuse()`
I added two variants, one which is unbalanced (always fusing a single arena into the group) while the other is balanced (always fusing two groups of equal size).
Unfortunately it is difficult to benchmark `upb_Arena_Free()` separately from fuse, since the only way to get a non-zero refcount is to fuse.
Results on my machine:
```
CPU: Intel Skylake Xeon with HyperThreading (48 cores) dL1:32KB dL2:1024KB dL3:38MB
Benchmark Time(ns) CPU(ns) Iterations
-------------------------------------------------------------------
BM_ArenaFuseUnbalanced/2 52.2 52.2 1000000 38.300M items/s
BM_ArenaFuseUnbalanced/8 416 416 142315 19.243M items/s
BM_ArenaFuseUnbalanced/64 4034 4033 14306 15.868M items/s
BM_ArenaFuseUnbalanced/128 8302 8301 6837 15.420M items/s
BM_ArenaFuseBalanced/2 54.7 54.7 1000000 36.581M items/s
BM_ArenaFuseBalanced/8 425 425 100000 18.845M items/s
BM_ArenaFuseBalanced/64 4029 4029 14632 15.886M items/s
BM_ArenaFuseBalanced/128 8050 8049 7176 15.902M items/s
```
PiperOrigin-RevId: 518292073
diff --git a/benchmarks/BUILD b/benchmarks/BUILD
index 07207fc..e219c42 100644
--- a/benchmarks/BUILD
+++ b/benchmarks/BUILD
@@ -87,7 +87,9 @@
":benchmark_descriptor_sv_cc_proto",
":benchmark_descriptor_upb_proto",
":benchmark_descriptor_upb_proto_reflection",
+ "//:base",
"//:descriptor_upb_proto",
+ "//:mem",
"//:reflection",
"@com_github_google_benchmark//:benchmark_main",
"@com_google_absl//absl/container:flat_hash_set",
diff --git a/benchmarks/benchmark.cc b/benchmarks/benchmark.cc
index 8ad30f0..58e76f8 100644
--- a/benchmarks/benchmark.cc
+++ b/benchmarks/benchmark.cc
@@ -35,6 +35,8 @@
#include "benchmarks/descriptor.upb.h"
#include "benchmarks/descriptor.upbdefs.h"
#include "benchmarks/descriptor_sv.pb.h"
+#include "upb/base/log2.h"
+#include "upb/mem/arena.h"
#include "upb/reflection/def.hpp"
upb_StringView descriptor = benchmarks_descriptor_proto_upbdefinit.descriptor;
@@ -73,6 +75,52 @@
}
BENCHMARK(BM_ArenaInitialBlockOneAlloc);
+static void BM_ArenaFuseUnbalanced(benchmark::State& state) {
+ std::vector<upb_Arena*> arenas(state.range(0));
+ size_t n = 0;
+ for (auto _ : state) {
+ for (auto& arena : arenas) {
+ arena = upb_Arena_New();
+ }
+ for (auto& arena : arenas) {
+ upb_Arena_Fuse(arenas[0], arena);
+ }
+ for (auto& arena : arenas) {
+ upb_Arena_Free(arena);
+ }
+ n += arenas.size();
+ }
+ state.SetItemsProcessed(n);
+}
+BENCHMARK(BM_ArenaFuseUnbalanced)->Range(2, 128);
+
+static void BM_ArenaFuseBalanced(benchmark::State& state) {
+ std::vector<upb_Arena*> arenas(state.range(0));
+ size_t n = 0;
+
+ for (auto _ : state) {
+ for (auto& arena : arenas) {
+ arena = upb_Arena_New();
+ }
+
+ // Perform a series of fuses that keeps the halves balanced.
+ size_t max = upb_Log2Ceiling(arenas.size());
+ for (size_t n = 0; n <= max; n++) {
+ size_t step = 1 << n;
+ for (size_t i = 0; i + step < arenas.size(); i += (step * 2)) {
+ upb_Arena_Fuse(arenas[i], arenas[i + step]);
+ }
+ }
+
+ for (auto& arena : arenas) {
+ upb_Arena_Free(arena);
+ }
+ n += arenas.size();
+ }
+ state.SetItemsProcessed(n);
+}
+BENCHMARK(BM_ArenaFuseBalanced)->Range(2, 128);
+
enum LoadDescriptorMode {
NoLayout,
WithLayout,
