Add single threaded thoughput optimized executor with high and low watermark at 25% / 75%.

3 files changed, 17 insertions, 4 deletions

diff --git a/searchlib/src/tests/common/sequencedtaskexecutor/sequencedtaskexecutor_benchmark.cpp b/searchlib/src/tests/common/sequencedtaskexecutor/sequencedtaskexecutor_benchmark.cpp
index b2b15ded274..2708399f653 100644
--- a/searchlib/src/tests/common/sequencedtaskexecutor/sequencedtaskexecutor_benchmark.cpp
+++ b/searchlib/src/tests/common/sequencedtaskexecutor/sequencedtaskexecutor_benchmark.cpp
@@ -10,13 +10,19 @@ using ExecutorId = search::ISequencedTaskExecutor::ExecutorId;
 int main(int argc, char *argv[]) {
     unsigned long numTasks = 1000000;
     unsigned numThreads = 4;
+    unsigned taskLimit = 1000;
+    SequencedTaskExecutor::Optimize optimize = SequencedTaskExecutor::Optimize::LATENCY;
     std::atomic<long> counter(0);
     if (argc > 1)
         numTasks = atol(argv[1]);
     if (argc > 2)
         numThreads = atoi(argv[2]);
+    if (argc > 3)
+        taskLimit = atoi(argv[3]);
+    if (argc > 4)
+        optimize = SequencedTaskExecutor::Optimize::THROUGHPUT;
 
-    auto executor = SequencedTaskExecutor::create(numThreads);
+    auto executor = SequencedTaskExecutor::create(numThreads, taskLimit, optimize);
     for (unsigned long tid(0); tid < numTasks; tid++) {
         executor->executeTask(ExecutorId(tid%numThreads), vespalib::makeLambdaTask([&counter] { counter++; }));
     }
diff --git a/searchlib/src/vespa/searchlib/common/sequencedtaskexecutor.cpp b/searchlib/src/vespa/searchlib/common/sequencedtaskexecutor.cpp
index 30723a6eb2a..32b515c8efc 100644
--- a/searchlib/src/vespa/searchlib/common/sequencedtaskexecutor.cpp
+++ b/searchlib/src/vespa/searchlib/common/sequencedtaskexecutor.cpp
@@ -2,8 +2,10 @@
 
 #include "sequencedtaskexecutor.h"
 #include <vespa/vespalib/util/blockingthreadstackexecutor.h>
+#include <vespa/vespalib/util/singleexecutor.h>
 
 using vespalib::BlockingThreadStackExecutor;
+using vespalib::SingleExecutor;
 
 namespace search {
 
@@ -15,12 +17,16 @@ constexpr uint32_t stackSize = 128 * 1024;
 
 
 std::unique_ptr<ISequencedTaskExecutor>
-SequencedTaskExecutor::create(uint32_t threads, uint32_t taskLimit)
+SequencedTaskExecutor::create(uint32_t threads, uint32_t taskLimit, Optimize optimize)
 {
     auto executors = std::make_unique<std::vector<std::unique_ptr<vespalib::SyncableThreadExecutor>>>();
     executors->reserve(threads);
     for (uint32_t id = 0; id < threads; ++id) {
-        executors->push_back(std::make_unique<BlockingThreadStackExecutor>(1, stackSize, taskLimit));
+        if (optimize == Optimize::THROUGHPUT) {
+            executors->push_back(std::make_unique<SingleExecutor>(taskLimit));
+        } else {
+            executors->push_back(std::make_unique<BlockingThreadStackExecutor>(1, stackSize, taskLimit));
+        }
     }
     return std::unique_ptr<ISequencedTaskExecutor>(new SequencedTaskExecutor(std::move(executors)));
 }
diff --git a/searchlib/src/vespa/searchlib/common/sequencedtaskexecutor.h b/searchlib/src/vespa/searchlib/common/sequencedtaskexecutor.h
index a29e3d5226c..a7a223dd23a 100644
--- a/searchlib/src/vespa/searchlib/common/sequencedtaskexecutor.h
+++ b/searchlib/src/vespa/searchlib/common/sequencedtaskexecutor.h
@@ -22,6 +22,7 @@ class SequencedTaskExecutor final : public ISequencedTaskExecutor
 
     SequencedTaskExecutor(std::unique_ptr<std::vector<std::unique_ptr<vespalib::SyncableThreadExecutor>>> executor);
 public:
+    enum class Optimize {LATENCY, THROUGHPUT};
     using ISequencedTaskExecutor::getExecutorId;
 
     ~SequencedTaskExecutor();
@@ -30,7 +31,7 @@ public:
     void executeTask(ExecutorId id, vespalib::Executor::Task::UP task) override;
     void sync() override;
     Stats getStats() override;
-    static std::unique_ptr<ISequencedTaskExecutor> create(uint32_t threads, uint32_t taskLimit = 1000);
+    static std::unique_ptr<ISequencedTaskExecutor> create(uint32_t threads, uint32_t taskLimit = 1000, Optimize optimize = Optimize::THROUGHPUT);
 };
 
 } // namespace search