Introduce memory usage estimation mode in data_frame_analyzer (elastic#584)

Author: przemekwitek

bin/data_frame_analyzer/CCmdLineParser.cc

@@ -20,6 +20,7 @@ const std::string CCmdLineParser::DESCRIPTION = "Usage: data_frame_analyzer [opt
 bool CCmdLineParser::parse(int argc,
                            const char* const* argv,
                            std::string& configFile,
+                           bool& memoryUsageEstimationOnly,
                            std::string& logProperties,
                            std::string& logPipe,
                            bool& lengthEncodedInput,
@@ -35,6 +36,7 @@ bool CCmdLineParser::parse(int argc,
             ("version", "Display version information and exit")
             ("config", boost::program_options::value<std::string>(),
                     "The configuration file")
+            ("memoryUsageEstimationOnly", "Whether to perform memory usage estimation only")
             ("logProperties", boost::program_options::value<std::string>(),
                     "Optional logger properties file")
             ("logPipe", boost::program_options::value<std::string>(),
@@ -66,6 +68,9 @@ bool CCmdLineParser::parse(int argc,
         if (vm.count("config") > 0) {
             configFile = vm["config"].as<std::string>();
         }
+        if (vm.count("memoryUsageEstimationOnly") > 0) {
+            memoryUsageEstimationOnly = true;
+        }
         if (vm.count("logProperties") > 0) {
             logProperties = vm["logProperties"].as<std::string>();
         }

bin/data_frame_analyzer/CCmdLineParser.h

@@ -27,6 +27,7 @@ class CCmdLineParser {
     static bool parse(int argc,
                       const char* const* argv,
                       std::string& configFile,
+                      bool& memoryUsageEstimationOnly,
                       std::string& logProperties,
                       std::string& logPipe,
                       bool& lengthEncodedInput,

bin/data_frame_analyzer/Main.cc

@@ -30,6 +30,7 @@
 #include <api/CDataFrameOutliersRunner.h>
 #include <api/CIoManager.h>
 #include <api/CLengthEncodedInputParser.h>
+#include <api/CMemoryUsageEstimationResultJsonWriter.h>
 
 #include "CCmdLineParser.h"
 
@@ -86,6 +87,7 @@ int main(int argc, char** argv) {
 
     // Read command line options
     std::string configFile;
+    bool memoryUsageEstimationOnly(false);
     std::string logProperties;
     std::string logPipe;
     bool lengthEncodedInput(false);
@@ -94,8 +96,9 @@ int main(int argc, char** argv) {
     std::string outputFileName;
     bool isOutputFileNamedPipe(false);
     if (ml::data_frame_analyzer::CCmdLineParser::parse(
-            argc, argv, configFile, logProperties, logPipe, lengthEncodedInput, inputFileName,
-            isInputFileNamedPipe, outputFileName, isOutputFileNamedPipe) == false) {
+            argc, argv, configFile, memoryUsageEstimationOnly, logProperties,
+            logPipe, lengthEncodedInput, inputFileName, isInputFileNamedPipe,
+            outputFileName, isOutputFileNamedPipe) == false) {
         return EXIT_FAILURE;
     }
 
@@ -127,13 +130,6 @@ int main(int argc, char** argv) {
     }
 
     using TInputParserUPtr = std::unique_ptr<ml::api::CInputParser>;
-    auto inputParser{[lengthEncodedInput, &ioMgr]() -> TInputParserUPtr {
-        if (lengthEncodedInput) {
-            return std::make_unique<ml::api::CLengthEncodedInputParser>(ioMgr.inputStream());
-        }
-        return std::make_unique<ml::api::CCsvInputParser>(
-            ioMgr.inputStream(), ml::api::CCsvInputParser::COMMA);
-    }()};
 
     std::string analysisSpecificationJson;
     bool couldReadConfigFile;
@@ -145,6 +141,16 @@ int main(int argc, char** argv) {
 
     auto analysisSpecification =
         std::make_unique<ml::api::CDataFrameAnalysisSpecification>(analysisSpecificationJson);
+
+    if (memoryUsageEstimationOnly) {
+        auto outStream = [&ioMgr]() {
+            return std::make_unique<ml::core::CJsonOutputStreamWrapper>(ioMgr.outputStream());
+        }();
+        ml::api::CMemoryUsageEstimationResultJsonWriter writer(*outStream);
+        analysisSpecification->estimateMemoryUsage(writer);
+        return EXIT_SUCCESS;
+    }
+
     if (analysisSpecification->numberThreads() > 1) {
         ml::core::startDefaultAsyncExecutor(analysisSpecification->numberThreads());
     }
@@ -156,6 +162,13 @@ int main(int argc, char** argv) {
 
     CCleanUpOnExit::add(dataFrameAnalyzer.dataFrameDirectory());
 
+    auto inputParser{[lengthEncodedInput, &ioMgr]() -> TInputParserUPtr {
+        if (lengthEncodedInput) {
+            return std::make_unique<ml::api::CLengthEncodedInputParser>(ioMgr.inputStream());
+        }
+        return std::make_unique<ml::api::CCsvInputParser>(
+            ioMgr.inputStream(), ml::api::CCsvInputParser::COMMA);
+    }()};
     if (inputParser->readStreamIntoVecs(
             [&dataFrameAnalyzer](const auto& fieldNames, const auto& fieldValues) {
                 return dataFrameAnalyzer.handleRecord(fieldNames, fieldValues);

include/api/CDataFrameAnalysisRunner.h

@@ -31,6 +31,7 @@ class CRowRef;
 }
 namespace api {
 class CDataFrameAnalysisSpecification;
+class CMemoryUsageEstimationResultJsonWriter;
 
 //! \brief Hierarchy for running a specific core::CDataFrame analyses.
 //!
@@ -75,6 +76,11 @@ class API_EXPORT CDataFrameAnalysisRunner {
     //! number of rows per subset.
     void computeAndSaveExecutionStrategy();
 
+    //! Estimates memory usage in two cases: one partition (the whole data frame
+    //! fits in main memory) and maximum tolerable number of partitions (only
+    //! one partition needs to be loaded to main memory).
+    void estimateMemoryUsage(CMemoryUsageEstimationResultJsonWriter& writer) const;
+
     //! Check if the data frame for this analysis should use in or out of core
     //! storage.
     bool storeDataFrameInMainMemory() const;

include/api/CDataFrameAnalysisSpecification.h

@@ -151,6 +151,11 @@ class API_EXPORT CDataFrameAnalysisSpecification {
     //! calling thread until the runner has finished.
     CDataFrameAnalysisRunner* run(const TStrVec& featureNames, core::CDataFrame& frame) const;
 
+    //! Estimates memory usage in two cases: one partition (the whole data frame
+    //! fits in main memory) and maximum tolerable number of partitions (only
+    //! one partition needs to be loaded to main memory).
+    void estimateMemoryUsage(CMemoryUsageEstimationResultJsonWriter& writer) const;
+
 private:
     void initializeRunner(const rapidjson::Value& jsonAnalysis);
 

include/api/CMemoryUsageEstimationResultJsonWriter.h

@@ -0,0 +1,42 @@
+/*
+ * Copyright Elasticsearch B.V. and/or licensed to Elasticsearch B.V. under one
+ * or more contributor license agreements. Licensed under the Elastic License;
+ * you may not use this file except in compliance with the Elastic License.
+ */
+#ifndef INCLUDED_ml_api_CMemoryUsageEstimationResultJsonWriter_h
+#define INCLUDED_ml_api_CMemoryUsageEstimationResultJsonWriter_h
+
+#include <core/CJsonOutputStreamWrapper.h>
+#include <core/CNonCopyable.h>
+#include <core/CRapidJsonConcurrentLineWriter.h>
+
+#include <api/ImportExport.h>
+
+#include <string>
+
+namespace ml {
+namespace api {
+
+//! \brief
+//! Write memory usage estimation result in JSON format
+//!
+//! DESCRIPTION:\n
+//! Outputs the memory usage estimation result.
+//!
+class API_EXPORT CMemoryUsageEstimationResultJsonWriter : private core::CNonCopyable {
+public:
+    //! \param[in] strmOut The wrapped stream to which to write output.
+    CMemoryUsageEstimationResultJsonWriter(core::CJsonOutputStreamWrapper& strmOut);
+
+    //! Writes the given memory usage estimation result in JSON format.
+    void write(const std::string& expectedMemoryUsageWithOnePartition,
+               const std::string& expectedMemoryUsageWithMaxPartitions);
+
+private:
+    //! JSON line writer
+    core::CRapidJsonConcurrentLineWriter m_Writer;
+};
+}
+}
+
+#endif // INCLUDED_ml_api_CMemoryUsageEstimationResultJsonWriter_h

lib/api/CDataFrameAnalysisRunner.cc

@@ -11,6 +11,7 @@
 #include <core/CScopedFastLock.h>
 
 #include <api/CDataFrameAnalysisSpecification.h>
+#include <api/CMemoryUsageEstimationResultJsonWriter.h>
 
 #include <boost/iterator/counting_iterator.hpp>
 
@@ -24,6 +25,13 @@ std::size_t memoryLimitWithSafetyMargin(const CDataFrameAnalysisSpecification& s
     return static_cast<std::size_t>(0.9 * static_cast<double>(spec.memoryLimit()) + 0.5);
 }
 
+std::size_t maximumNumberPartitions(const CDataFrameAnalysisSpecification& spec) {
+    // We limit the maximum number of partitions to rows^(1/2) because very
+    // large numbers of partitions are going to be slow and it is better to tell
+    // user to allocate more resources for the job in this case.
+    return static_cast<std::size_t>(std::sqrt(static_cast<double>(spec.numberRows())) + 0.5);
+}
+
 const std::size_t MAXIMUM_FRACTIONAL_PROGRESS{std::size_t{1}
                                               << ((sizeof(std::size_t) - 2) * 8)};
 }
@@ -36,6 +44,25 @@ CDataFrameAnalysisRunner::~CDataFrameAnalysisRunner() {
     this->waitToFinish();
 }
 
+void CDataFrameAnalysisRunner::estimateMemoryUsage(CMemoryUsageEstimationResultJsonWriter& writer) const {
+    std::size_t numberRows{m_Spec.numberRows()};
+    std::size_t numberColumns{m_Spec.numberColumns() + this->numberExtraColumns()};
+    std::size_t maxNumberPartitions{maximumNumberPartitions(m_Spec)};
+    if (maxNumberPartitions == 0) {
+        writer.write("0", "0");
+        return;
+    }
+    std::size_t expectedMemoryUsageWithOnePartition{
+        this->estimateMemoryUsage(numberRows, numberRows, numberColumns)};
+    std::size_t expectedMemoryUsageWithMaxPartitions{this->estimateMemoryUsage(
+        numberRows, numberRows / maxNumberPartitions, numberColumns)};
+    auto roundUpToNearestKilobyte = [](std::size_t bytes) {
+        return std::to_string((bytes + 1024 - 1) / 1024) + "kB";
+    };
+    writer.write(roundUpToNearestKilobyte(expectedMemoryUsageWithOnePartition),
+                 roundUpToNearestKilobyte(expectedMemoryUsageWithMaxPartitions));
+}
+
 void CDataFrameAnalysisRunner::computeAndSaveExecutionStrategy() {
 
     std::size_t numberRows{m_Spec.numberRows()};
@@ -45,17 +72,12 @@ void CDataFrameAnalysisRunner::computeAndSaveExecutionStrategy() {
     LOG_TRACE(<< "memory limit = " << memoryLimit);
 
     // Find the smallest number of partitions such that the size per partition
-    // is less than the memory limit. We limit this to rows^(1/2) because very
-    // large numbers of partitions are going to be slow and it is better to tell
-    // user to allocate more resources for the job in this case.
-
-    std::size_t maximumNumberPartitions{
-        static_cast<std::size_t>(std::sqrt(static_cast<double>(numberRows)) + 0.5)};
+    // is less than the memory limit.
 
+    std::size_t maxNumberPartitions{maximumNumberPartitions(m_Spec)};
     std::size_t memoryUsage{0};
 
-    for (m_NumberPartitions = 1; m_NumberPartitions < maximumNumberPartitions;
-         ++m_NumberPartitions) {
+    for (m_NumberPartitions = 1; m_NumberPartitions < maxNumberPartitions; ++m_NumberPartitions) {
         std::size_t partitionNumberRows{numberRows / m_NumberPartitions};
         memoryUsage = this->estimateMemoryUsage(numberRows, partitionNumberRows, numberColumns);
         LOG_TRACE(<< "partition number rows = " << partitionNumberRows);

lib/api/CDataFrameAnalysisSpecification.cc

@@ -12,6 +12,7 @@
 #include <api/CDataFrameAnalysisConfigReader.h>
 #include <api/CDataFrameBoostedTreeRunner.h>
 #include <api/CDataFrameOutliersRunner.h>
+#include <api/CMemoryUsageEstimationResultJsonWriter.h>
 
 #include <rapidjson/document.h>
 #include <rapidjson/ostreamwrapper.h>
@@ -184,6 +185,14 @@ CDataFrameAnalysisRunner* CDataFrameAnalysisSpecification::run(const TStrVec& fe
     return nullptr;
 }
 
+void CDataFrameAnalysisSpecification::estimateMemoryUsage(CMemoryUsageEstimationResultJsonWriter& writer) const {
+    if (m_Runner == nullptr) {
+        HANDLE_FATAL(<< "Internal error: no runner available so can't estimate memory. Please report this problem.");
+        return;
+    }
+    m_Runner->estimateMemoryUsage(writer);
+}
+
 void CDataFrameAnalysisSpecification::initializeRunner(const rapidjson::Value& jsonAnalysis) {
     // We pass of the interpretation of the parameters object to the appropriate
     // analysis runner.

lib/api/CMemoryUsageEstimationResultJsonWriter.cc

@@ -0,0 +1,35 @@
+/*
+ * Copyright Elasticsearch B.V. and/or licensed to Elasticsearch B.V. under one
+ * or more contributor license agreements. Licensed under the Elastic License;
+ * you may not use this file except in compliance with the Elastic License.
+ */
+
+#include <api/CMemoryUsageEstimationResultJsonWriter.h>
+
+namespace ml {
+namespace api {
+namespace {
+
+// JSON field names
+const std::string EXPECTED_MEMORY_USAGE_WITH_ONE_PARTITION("expected_memory_usage_with_one_partition");
+const std::string EXPECTED_MEMORY_USAGE_WITH_MAX_PARTITIONS("expected_memory_usage_with_max_partitions");
+}
+
+CMemoryUsageEstimationResultJsonWriter::CMemoryUsageEstimationResultJsonWriter(core::CJsonOutputStreamWrapper& strmOut)
+    : m_Writer(strmOut) {
+    // Don't write any output in the constructor because, the way things work at
+    // the moment, the output stream might be redirected after construction
+}
+
+void CMemoryUsageEstimationResultJsonWriter::write(const std::string& expectedMemoryUsageWithOnePartition,
+                                                   const std::string& expectedMemoryUsageWithMaxPartitions) {
+    m_Writer.StartObject();
+    m_Writer.Key(EXPECTED_MEMORY_USAGE_WITH_ONE_PARTITION);
+    m_Writer.String(expectedMemoryUsageWithOnePartition);
+    m_Writer.Key(EXPECTED_MEMORY_USAGE_WITH_MAX_PARTITIONS);
+    m_Writer.String(expectedMemoryUsageWithMaxPartitions);
+    m_Writer.EndObject();
+    m_Writer.flush();
+}
+}
+}

lib/api/Makefile.first

@@ -43,6 +43,7 @@ CInputParser.cc \
 CIoManager.cc \
 CJsonOutputWriter.cc \
 CLengthEncodedInputParser.cc \
+CMemoryUsageEstimationResultJsonWriter.cc \
 CModelPlotDataJsonWriter.cc \
 CModelSizeStatsJsonWriter.cc \
 CModelSnapshotJsonWriter.cc \