From 3c3555da7adc307044715d5382c254c2c7997bce Mon Sep 17 00:00:00 2001
From: Andrew Noyes <andrew@weaselab.dev>
Date: Tue, 26 Aug 2025 13:33:16 -0400
Subject: [PATCH] Benchmark different batches and wait strategies

---
 benchmarks/bench_thread_pipeline.cpp | 166 +++++++++++++++++++++++++++
 1 file changed, 166 insertions(+)

diff --git a/benchmarks/bench_thread_pipeline.cpp b/benchmarks/bench_thread_pipeline.cpp
index c77115a..2e643d3 100644
--- a/benchmarks/bench_thread_pipeline.cpp
+++ b/benchmarks/bench_thread_pipeline.cpp
@@ -80,6 +80,172 @@ int main() {
     stage0_thread.join();
   }
 
+  // Batch size comparison benchmark
+  {
+    constexpr int LOG_PIPELINE_SIZE = 10; // 2^10 = 1024 slots
+    constexpr int NUM_ITEMS = 100'000;
+    constexpr int BUSY_ITERS = 100;
+
+    auto bench = ankerl::nanobench::Bench()
+                     .title("Batch Size Impact")
+                     .unit("item")
+                     .batch(NUM_ITEMS)
+                     .relative(true)
+                     .warmup(100);
+
+    for (int batch_size : {1, 4, 16, 64, 256}) {
+      std::vector<int> threads_per_stage = {1};
+      ThreadPipeline<std::latch *> pipeline(LOG_PIPELINE_SIZE,
+                                            threads_per_stage);
+
+      std::latch done{0};
+
+      // Stage 0 consumer thread
+      std::thread stage0_thread([&pipeline, &done]() {
+        const int stage = 0;
+        const int thread_id = 0;
+
+        for (;;) {
+          auto guard = pipeline.acquire(stage, thread_id);
+
+          for (auto &item : guard.batch) {
+            for (volatile int i = 0; i < BUSY_ITERS; i = i + 1) {
+            }
+            if (item == &done) {
+              return;
+            }
+            if (item) {
+              item->count_down();
+            }
+          }
+        }
+      });
+
+      bench.run("Batch size " + std::to_string(batch_size), [&] {
+        // Producer (main thread)
+        int items_pushed = 0;
+        while (items_pushed < NUM_ITEMS - 1) {
+          auto guard = pipeline.push(
+              std::min(NUM_ITEMS - 1 - items_pushed, batch_size), true);
+
+          auto it = guard.batch.begin();
+          items_pushed += guard.batch.size();
+          for (size_t i = 0; i < guard.batch.size(); ++i, ++it) {
+            *it = nullptr;
+          }
+        }
+        std::latch finish{1};
+        {
+          auto guard = pipeline.push(1, true);
+          *guard.batch.begin() = &finish;
+        }
+        finish.wait();
+      });
+
+      {
+        auto guard = pipeline.push(1, true);
+        *guard.batch.begin() = &done;
+      }
+
+      stage0_thread.join();
+    }
+  }
+
+  // Helper function for wait strategy benchmarks
+  auto benchmark_wait_strategy =
+      []<WaitStrategy strategy>(const std::string &name,
+                                ankerl::nanobench::Bench &bench) {
+        constexpr int LOG_PIPELINE_SIZE =
+            8; // Smaller buffer to increase contention
+        constexpr int NUM_ITEMS = 50'000;
+        constexpr int BATCH_SIZE = 4; // Small batches to increase coordination
+        constexpr int BUSY_ITERS =
+            10; // Light work to emphasize coordination overhead
+
+        std::vector<int> threads_per_stage = {1, 1}; // Two stages
+        ThreadPipeline<std::latch *, strategy> pipeline(LOG_PIPELINE_SIZE,
+                                                        threads_per_stage);
+
+        std::latch done{0};
+
+        // Stage 0 worker
+        std::thread stage0_thread([&pipeline, &done]() {
+          const int stage = 0;
+          const int thread_id = 0;
+          for (;;) {
+            auto guard = pipeline.acquire(stage, thread_id);
+            for (auto &item : guard.batch) {
+              for (volatile int i = 0; i < BUSY_ITERS; i = i + 1) {
+              }
+              if (item == &done)
+                return;
+            }
+          }
+        });
+
+        // Stage 1 worker (final stage - always calls futex wake)
+        std::thread stage1_thread([&pipeline, &done]() {
+          const int stage = 1;
+          const int thread_id = 0;
+          for (;;) {
+            auto guard = pipeline.acquire(stage, thread_id);
+            for (auto &item : guard.batch) {
+              for (volatile int i = 0; i < BUSY_ITERS; i = i + 1) {
+              }
+              if (item == &done)
+                return;
+              if (item)
+                item->count_down();
+            }
+          }
+        });
+
+        bench.run(name, [&] {
+          int items_pushed = 0;
+          while (items_pushed < NUM_ITEMS - 1) {
+            auto guard = pipeline.push(
+                std::min(NUM_ITEMS - 1 - items_pushed, BATCH_SIZE), true);
+            auto it = guard.batch.begin();
+            items_pushed += guard.batch.size();
+            for (size_t i = 0; i < guard.batch.size(); ++i, ++it) {
+              *it = nullptr;
+            }
+          }
+          std::latch finish{1};
+          {
+            auto guard = pipeline.push(1, true);
+            *guard.batch.begin() = &finish;
+          }
+          finish.wait();
+        });
+
+        // Shutdown
+        {
+          auto guard = pipeline.push(1, true);
+          auto it = guard.batch.begin();
+          *it = &done;
+        }
+        stage0_thread.join();
+        stage1_thread.join();
+      };
+
+  // Wait strategy comparison benchmark - multiple stages to trigger futex wakes
+  {
+    auto bench = ankerl::nanobench::Bench()
+                     .title("Wait Strategy Comparison")
+                     .unit("item")
+                     .batch(50'000)
+                     .relative(true)
+                     .warmup(50);
+
+    benchmark_wait_strategy.template operator()<WaitStrategy::WaitIfStageEmpty>(
+        "WaitIfStageEmpty", bench);
+    benchmark_wait_strategy.template
+    operator()<WaitStrategy::WaitIfUpstreamIdle>("WaitIfUpstreamIdle", bench);
+    benchmark_wait_strategy.template operator()<WaitStrategy::Never>("Never",
+                                                                     bench);
+  }
+
   // TODO: Add more benchmarks for:
   // - Multi-stage pipelines (3+ stages)
   // - Multiple threads per stage