In order to reduce overhead we look at the hit estimate before fanning out to multiple threads.

Default is as today, configurable per rank-profile via min-hits-per-thread.
Default should be changed to a sane number once verified.
The main intentions is to counter the sometimes high cost of per thread rank-setup.

7 files changed, 36 insertions, 2 deletions

diff --git a/searchlib/src/tests/fef/properties/properties_test.cpp b/searchlib/src/tests/fef/properties/properties_test.cpp
index a08d511b418..a904fd13e49 100644
--- a/searchlib/src/tests/fef/properties/properties_test.cpp
+++ b/searchlib/src/tests/fef/properties/properties_test.cpp
@@ -267,6 +267,15 @@ TEST("test stuff") {
             p.add("vespa.matching.numthreadspersearch", "50");
             EXPECT_EQUAL(matching::NumThreadsPerSearch::lookup(p), 50u);
         }
+
+        { // vespa.matching.minhitsperthread
+            EXPECT_EQUAL(matching::MinHitsPerThread::NAME, vespalib::string("vespa.matching.minhitsperthread"));
+            EXPECT_EQUAL(matching::MinHitsPerThread::DEFAULT_VALUE, 0);
+            Properties p;
+            EXPECT_EQUAL(matching::MinHitsPerThread::lookup(p), 0);
+            p.add("vespa.matching.minhitsperthread", "50");
+            EXPECT_EQUAL(matching::MinHitsPerThread::lookup(p), 50u);
+        }
         {
             EXPECT_EQUAL(matching::NumSearchPartitions::NAME, vespalib::string("vespa.matching.numsearchpartitions"));
             EXPECT_EQUAL(matching::NumSearchPartitions::DEFAULT_VALUE, 1u);
diff --git a/searchlib/src/tests/ranksetup/ranksetup_test.cpp b/searchlib/src/tests/ranksetup/ranksetup_test.cpp
index aee04ef4cb7..ead2f09766d 100644
--- a/searchlib/src/tests/ranksetup/ranksetup_test.cpp
+++ b/searchlib/src/tests/ranksetup/ranksetup_test.cpp
@@ -485,6 +485,7 @@ void RankSetupTest::testRankSetup()
     env.getProperties().add(dump::Feature::NAME, "foo");
     env.getProperties().add(dump::Feature::NAME, "bar");
     env.getProperties().add(matching::NumThreadsPerSearch::NAME, "3");
+    env.getProperties().add(matching::MinHitsPerThread::NAME, "8");
     env.getProperties().add(matchphase::DegradationAttribute::NAME, "mystaticrankattr");
     env.getProperties().add(matchphase::DegradationAscendingOrder::NAME, "true");
     env.getProperties().add(matchphase::DegradationMaxHits::NAME, "12345");
@@ -509,6 +510,7 @@ void RankSetupTest::testRankSetup()
     EXPECT_EQUAL(rs.getDumpFeatures()[0], vespalib::string("foo"));
     EXPECT_EQUAL(rs.getDumpFeatures()[1], vespalib::string("bar"));
     EXPECT_EQUAL(rs.getNumThreadsPerSearch(), 3u);
+    EXPECT_EQUAL(rs.getMinHitsPerThread(), 8u);
     EXPECT_EQUAL(rs.getDegradationAttribute(), "mystaticrankattr");
     EXPECT_EQUAL(rs.isDegradationOrderAscending(), true);
     EXPECT_EQUAL(rs.getDegradationMaxHits(), 12345u);
diff --git a/searchlib/src/vespa/searchlib/fef/indexproperties.cpp b/searchlib/src/vespa/searchlib/fef/indexproperties.cpp
index 14d92e3a909..40d2b0ba811 100644
--- a/searchlib/src/vespa/searchlib/fef/indexproperties.cpp
+++ b/searchlib/src/vespa/searchlib/fef/indexproperties.cpp
@@ -161,6 +161,17 @@ NumSearchPartitions::lookup(const Properties &props)
 {
     return lookupUint32(props, NAME, DEFAULT_VALUE);
 }
+
+const vespalib::string MinHitsPerThread::NAME("vespa.matching.minhitsperthread");
+const uint32_t MinHitsPerThread::DEFAULT_VALUE(0);
+
+uint32_t
+MinHitsPerThread::lookup(const Properties &props)
+{
+    return lookupUint32(props, NAME, DEFAULT_VALUE);
+}
+
+
 } // namespace matching
 
 namespace matchphase {
diff --git a/searchlib/src/vespa/searchlib/fef/indexproperties.h b/searchlib/src/vespa/searchlib/fef/indexproperties.h
index 8dcd08dfc49..626206836ca 100644
--- a/searchlib/src/vespa/searchlib/fef/indexproperties.h
+++ b/searchlib/src/vespa/searchlib/fef/indexproperties.h
@@ -105,6 +105,14 @@ namespace matching {
     /**
      * Property for the number of threads used per search.
      **/
+    struct MinHitsPerThread {
+        static const vespalib::string NAME;
+        static const uint32_t DEFAULT_VALUE;
+        static uint32_t lookup(const Properties &props);
+    };
+    /**
+     * Property for the number of threads used per search.
+     **/
     struct NumSearchPartitions {
         static const vespalib::string NAME;
         static const uint32_t DEFAULT_VALUE;
diff --git a/searchlib/src/vespa/searchlib/fef/ranksetup.cpp b/searchlib/src/vespa/searchlib/fef/ranksetup.cpp
index a954f70c82b..bfc79482080 100644
--- a/searchlib/src/vespa/searchlib/fef/ranksetup.cpp
+++ b/searchlib/src/vespa/searchlib/fef/ranksetup.cpp
@@ -36,6 +36,7 @@ RankSetup::RankSetup(const BlueprintFactory &factory,
       _secondPhaseRankFeature(),
       _degradationAttribute(),
       _numThreads(0),
+      _minHitsPerThread(0),
       _numSearchPartitions(0),
       _heapSize(0),
       _arraySize(0),
@@ -75,6 +76,7 @@ RankSetup::configure()
     }
     set_termwise_limit(indexproperties::matching::TermwiseLimit::lookup(_indexEnv.getProperties()));
     setNumThreadsPerSearch(indexproperties::matching::NumThreadsPerSearch::lookup(_indexEnv.getProperties()));
+    setMinHitsPerThread(indexproperties::matching::MinHitsPerThread::lookup(_indexEnv.getProperties()));
     setNumSearchPartitions(indexproperties::matching::NumSearchPartitions::lookup(_indexEnv.getProperties()));
     setHeapSize(indexproperties::hitcollector::HeapSize::lookup(_indexEnv.getProperties()));
     setArraySize(indexproperties::hitcollector::ArraySize::lookup(_indexEnv.getProperties()));
diff --git a/searchlib/src/vespa/searchlib/fef/ranksetup.h b/searchlib/src/vespa/searchlib/fef/ranksetup.h
index 86b381e3af6..d5acb13e117 100644
--- a/searchlib/src/vespa/searchlib/fef/ranksetup.h
+++ b/searchlib/src/vespa/searchlib/fef/ranksetup.h
@@ -34,6 +34,7 @@ private:
     vespalib::string         _degradationAttribute;
     double                   _termwise_limit;
     uint32_t                 _numThreads;
+    uint32_t                 _minHitsPerThread;
     uint32_t                 _numSearchPartitions;
     uint32_t                 _heapSize;
     uint32_t                 _arraySize;
@@ -146,6 +147,8 @@ public:
      * @return the number of threads
      **/
     uint32_t getNumThreadsPerSearch() const { return _numThreads; }
+    uint32_t getMinHitsPerThread() const { return _minHitsPerThread; }
+    void setMinHitsPerThread(uint32_t minHitsPerThread) { _minHitsPerThread = minHitsPerThread; }
 
     void setNumSearchPartitions(uint32_t numSearchPartitions) { _numSearchPartitions = numSearchPartitions; }
 
diff --git a/searchlib/src/vespa/searchlib/queryeval/blueprint.h b/searchlib/src/vespa/searchlib/queryeval/blueprint.h
index 2700f218321..b01d60d529c 100644
--- a/searchlib/src/vespa/searchlib/queryeval/blueprint.h
+++ b/searchlib/src/vespa/searchlib/queryeval/blueprint.h
@@ -259,8 +259,7 @@ public:
     Blueprint::UP removeChild(size_t n);
     SearchIterator::UP createSearch(fef::MatchData &md, bool strict) const override;
 
-    virtual HitEstimate
-    combine(const std::vector<HitEstimate> &data) const = 0;
+    virtual HitEstimate combine(const std::vector<HitEstimate> &data) const = 0;
     virtual FieldSpecBaseList exposeFields() const = 0;
     virtual void sort(std::vector<Blueprint*> &children) const = 0;
     virtual bool inheritStrict(size_t i) const = 0;