Run prepare step of two-phase puts in the shared executor instead of the attribute field executor.

Benchmarking has shown that we get an uneven balance between the underlying threads when using the attribute field executor.
This is a sequenced task executor, where each underlying thread only is allowed to handle a subset of the prepare tasks (based on the local document id).
Using the shared executor should give an even balance.

1 files changed, 31 insertions, 0 deletions

diff --git a/staging_vespalib/src/vespa/vespalib/util/foreground_thread_executor.h b/staging_vespalib/src/vespa/vespalib/util/foreground_thread_executor.h
new file mode 100644
index 00000000000..575552971fa
--- /dev/null
+++ b/staging_vespalib/src/vespa/vespalib/util/foreground_thread_executor.h
@@ -0,0 +1,31 @@
+// Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+#pragma once
+
+#include <vespa/vespalib/util/threadexecutor.h>
+#include <atomic>
+
+namespace vespalib {
+
+/**
+ * Implementation of the ThreadExecutor interface that runs all tasks in the foreground by the calling thread.
+ */
+class ForegroundThreadExecutor : public vespalib::ThreadExecutor {
+private:
+    std::atomic<size_t> _accepted;
+
+public:
+    ForegroundThreadExecutor() : _accepted(0) { }
+    Task::UP execute(Task::UP task) override {
+        task->run();
+        ++_accepted;
+        return Task::UP();
+    }
+    size_t getNumThreads() const override { return 0; }
+    Stats getStats() override {
+        return ExecutorStats(ExecutorStats::QueueSizeT(), _accepted.load(std::memory_order_relaxed), 0);
+    }
+    virtual void setTaskLimit(uint32_t taskLimit) override { (void) taskLimit; }
+};
+
+}