tdf#98955 hardware_concurrency not ideal for thread pools

A new static member getPreferredConcurrency added to
comphelper::ThreadPool to return a configurable max
number of threads.

By default the new function returns the hardware_concurrency
value provided by std::thread. When MAX_CONCURRENCY envar is
defined, the return value is limited to whatever is set there.

Three call-sites that used std::thread::hardware_concurrency
have been replaced with getPreferredConcurrency.

Unittests added to cover the functionality of the new member.

Unittests are capped to 4 threads.

Change-Id: I3332e393a88a5ed436316fa712ed920a4b37f4af
Reviewed-on: https://gerrit.libreoffice.org/26254Tested-by: Jenkins <ci@libreoffice.org>
Reviewed-by: Ashod Nakashian <ashnakash@gmail.com>

comphelper/CppunitTest_comphelper_threadpool_test.mk

+# -*- Mode: makefile-gmake; tab-width: 4; indent-tabs-mode: t -*-
+#
+# This file is part of the LibreOffice project.
+#
+# This Source Code Form is subject to the terms of the Mozilla Public
+# License, v. 2.0. If a copy of the MPL was not distributed with this
+# file, You can obtain one at http://mozilla.org/MPL/2.0/.
+#
+
+$(eval $(call gb_CppunitTest_CppunitTest,comphelper_threadpool_test))
+
+$(eval $(call gb_CppunitTest_add_exception_objects,comphelper_threadpool_test, \
+    comphelper/qa/unit/threadpooltest \
+))
+
+$(eval $(call gb_CppunitTest_use_externals,comphelper_threadpool_test,\
+	boost_headers \
+))
+
+$(eval $(call gb_CppunitTest_use_sdk_api,comphelper_threadpool_test))
+
+$(eval $(call gb_CppunitTest_use_libraries,comphelper_threadpool_test, \
+    comphelper \
+    cppuhelper \
+    cppu \
+    sal \
+	$(gb_UWINAPI) \
+))
+
+# vim: set noet sw=4 ts=4:

comphelper/Module_comphelper.mk

@@ -28,9 +28,10 @@ $(eval $(call gb_Module_add_subsequentcheck_targets,comphelper,\
 ))
 
 $(eval $(call gb_Module_add_check_targets,comphelper,\
+    CppunitTest_comphelper_threadpool_test \
     CppunitTest_comphelper_syntaxhighlight_test \
     CppunitTest_comphelper_variadictemplates_test \
-	CppunitTest_comphelper_ifcontainer \
+    CppunitTest_comphelper_ifcontainer \
     CppunitTest_comphelper_test \
 ))
 

comphelper/qa/unit/threadpooltest.cxx

+/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4 -*- */
+/*
+ * This file is part of the LibreOffice project.
+ *
+ * This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/.
+ */
+
+#include <comphelper/threadpool.hxx>
+#include "cppunit/TestAssert.h"
+#include "cppunit/TestFixture.h"
+#include "cppunit/extensions/HelperMacros.h"
+#include "cppunit/plugin/TestPlugIn.h"
+
+#include <stdlib.h>
+#include <thread>
+
+class ThreadPoolTest : public CppUnit::TestFixture
+{
+public:
+    void testPreferredConcurrency();
+
+    CPPUNIT_TEST_SUITE(ThreadPoolTest);
+    CPPUNIT_TEST(testPreferredConcurrency);
+    CPPUNIT_TEST_SUITE_END();
+};
+
+void ThreadPoolTest::testPreferredConcurrency() {
+
+    // Check default.
+    auto nThreads = comphelper::ThreadPool::getPreferredConcurrency();
+    sal_Int32 nExpected = 4; // UTs are capped to 4.
+    CPPUNIT_ASSERT_EQUAL(nExpected, nThreads);
+
+#ifndef _WIN32_WINNT
+    // The result should be cached, so this should change anything.
+    nThreads = std::thread::hardware_concurrency() * 2;
+    setenv("MAX_CONCURRENCY", std::to_string(nThreads).c_str(), true);
+    nThreads = comphelper::ThreadPool::getPreferredConcurrency();
+    CPPUNIT_ASSERT_MESSAGE("Expected no more than hardware threads",
+                           nThreads <= (sal_Int32)std::thread::hardware_concurrency());
+
+    // Revert and check. Again, nothing should change.
+    unsetenv("MAX_CONCURRENCY");
+    nThreads = comphelper::ThreadPool::getPreferredConcurrency();
+    CPPUNIT_ASSERT_EQUAL(nExpected, nThreads);
+#endif
+}
+
+CPPUNIT_TEST_SUITE_REGISTRATION(ThreadPoolTest);
+
+CPPUNIT_PLUGIN_IMPLEMENT();
+
+/* vim:set shiftwidth=4 softtabstop=4 expandtab: */

comphelper/source/misc/threadpool.cxx

@@ -10,6 +10,7 @@
 #include <comphelper/threadpool.hxx>
 
 #include <rtl/instance.hxx>
+#include <rtl/string.hxx>
 #include <algorithm>
 #include <memory>
 #include <thread>
@@ -115,7 +116,7 @@ struct ThreadPoolStatic : public rtl::StaticWithInit< std::shared_ptr< ThreadPoo
                                                       ThreadPoolStatic >
 {
     std::shared_ptr< ThreadPool > operator () () {
-        sal_Int32 nThreads = std::max( std::thread::hardware_concurrency(), 1U );
+        const sal_Int32 nThreads = ThreadPool::getPreferredConcurrency();
         return std::make_shared< ThreadPool >( nThreads );
     };
 };
@@ -125,6 +126,27 @@ ThreadPool& ThreadPool::getSharedOptimalPool()
     return *ThreadPoolStatic::get().get();
 }
 
+sal_Int32 ThreadPool::getPreferredConcurrency()
+{
+    static sal_Int32 ThreadCount = 0;
+    if (ThreadCount == 0)
+    {
+        const sal_Int32 nHardThreads = std::max(std::thread::hardware_concurrency(), 1U);
+        sal_Int32 nThreads = nHardThreads;
+        const char *pEnv = getenv("MAX_CONCURRENCY");
+        if (pEnv != nullptr)
+        {
+            // Override with user/admin preferrence.
+            nThreads = rtl_str_toInt32(pEnv, 10);
+        }
+
+        nThreads = std::min(nHardThreads, nThreads);
+        ThreadCount = std::max<sal_Int32>(nThreads, 1);
+    }
+
+    return ThreadCount;
+}
+
 void ThreadPool::waitAndCleanupWorkers()
 {
     waitUntilEmpty();

include/comphelper/threadpool.hxx

@@ -36,6 +36,12 @@ public:
     /// count for the CPU
     static      ThreadPool& getSharedOptimalPool();
 
+    /// returns a configurable max-concurrency
+    /// limit to avoid spawning an unnecessarily
+    /// large number of threads on high-core boxes.
+    /// MAX_CONCURRENCY envar controls the cap.
+    static      sal_Int32 getPreferredConcurrency();
+
                 ThreadPool( sal_Int32 nWorkers );
     virtual    ~ThreadPool();
 

package/source/zippackage/ZipPackageStream.cxx

@@ -832,7 +832,7 @@ bool ZipPackageStream::saveChild(
                     // cores and allow 4-times the amount for having the queue well filled. The
                     // 2nd parameter is the time to wait between cleanups in 10th of a second.
                     // Both values may be added to the configuration settings if needed.
-                    static sal_Int32 nAllowedThreads(std::max(std::thread::hardware_concurrency(), 1U) * 4);
+                    static sal_Int32 nAllowedThreads(comphelper::ThreadPool::getPreferredConcurrency() * 4);
                     rZipOut.reduceScheduledThreadsToGivenNumberOrLess(nAllowedThreads, 1);
 
                     // Start a new thread deflating this zip entry

sc/source/filter/excel/xetable.cxx

@@ -2133,7 +2133,7 @@ void XclExpRowBuffer::Finalize( XclExpDefaultRowData& rDefRowData, const ScfUInt
     // This is staggeringly slow, and each element operates only
     // on its own data.
     const size_t nRows = maRowMap.size();
-    const size_t nThreads = nRows < 128 ? 1 : std::max(std::thread::hardware_concurrency(), 1U);
+    const size_t nThreads = nRows < 128 ? 1 : comphelper::ThreadPool::getPreferredConcurrency();
 #else
     const size_t nThreads = 1; // globally disable multi-threading for now.
 #endif

sc/source/filter/oox/workbookfragment.cxx

@@ -309,7 +309,7 @@ public:
 
 void importSheetFragments( WorkbookFragment& rWorkbookHandler, SheetFragmentVector& rSheets )
 {
-    sal_Int32 nThreads = std::min( rSheets.size(), (size_t) std::max(std::thread::hardware_concurrency(), 1U) );
+    sal_Int32 nThreads = std::min( rSheets.size(), (size_t) comphelper::ThreadPool::getPreferredConcurrency() );
 
     Reference< XComponentContext > xContext = comphelper::getProcessComponentContext();
 

solenv/gbuild/CppunitTest.mk

@@ -19,6 +19,9 @@
 
 # CppunitTest class
 
+# Cap the number of threads unittests use.
+export MAX_CONCURRENCY=4
+
 gb_CppunitTest_UNITTESTFAILED ?= $(GBUILDDIR)/platform/unittest-failed-default.sh
 gb_CppunitTest_PYTHONDEPS ?= $(call gb_Library_get_target,pyuno_wrapper) $(if $(SYSTEM_PYTHON),,$(call gb_Package_get_target,python3))