Growing virtual memory usage on multithreaded benchmark #806
Description
Hi, there! It looks like on multithreaded benchmarks virtual memory keeps growing and is never freed. Valgrind only shows a leak of 240 bytes due to #782 so not sure why virtual mem only keeps growing. Also, I only see this on multithreaded benchmarks.
For example, on this bench below, which just adds numbers to a vector<uint64_t>, virtual mem quickly gets to upwards of 60GB and never seems to get freed in between benchmark runs:
#include <vector>
#include "benchmark/benchmark.h"
namespace bench {
const int kBMRangeMultiplier = 10;
const int kBMRangeLower = 1000;
const int kBMRangeUpper = 500000000;
static void BM_MemGrowthMultithreaded(benchmark::State& state) {
std::vector<uint64_t> numbers;
numbers.reserve(state.range(0));
for (int i = 0; i < state.range(0); ++i) {
numbers.push_back(i);
}
for (auto _ : state) {
size_t size = state.range(0);
std::vector<uint64_t> results;
results.reserve(size);
// It doesn't matter what we're adding, just creating some load.
for (size_t i = 0; i < size; i++) {
results.push_back(i);
}
state.counters["matches"] =
benchmark::Counter(results.size(), benchmark::Counter::kAvgThreads);
}
state.SetComplexityN(state.range(0));
}
// Single-threaded.
BENCHMARK(BM_MemGrowthMultithreaded)
->Threads(1)
->UseRealTime()
->RangeMultiplier(kBMRangeMultiplier)
->Range(kBMRangeLower, kBMRangeUpper)
->Unit(benchmark::kMillisecond)
->Complexity(benchmark::oN);
// 2 threads.
BENCHMARK(BM_MemGrowthMultithreaded)
->Threads(2)
->UseRealTime()
->RangeMultiplier(kBMRangeMultiplier)
->Range(kBMRangeLower, kBMRangeUpper)
->Unit(benchmark::kMillisecond)
->Complexity(benchmark::oN);
// 4 threads.
BENCHMARK(BM_MemGrowthMultithreaded)
->Threads(4)
->UseRealTime()
->RangeMultiplier(kBMRangeMultiplier)
->Range(kBMRangeLower, kBMRangeUpper)
->Unit(benchmark::kMillisecond)
->Complexity(benchmark::oN);
// 8 threads.
BENCHMARK(BM_MemGrowthMultithreaded)
->Threads(8)
->UseRealTime()
->RangeMultiplier(kBMRangeMultiplier)
->Range(kBMRangeLower, kBMRangeUpper)
->Unit(benchmark::kMillisecond)
->Complexity(benchmark::oN);
} // namespace bench
BENCHMARK_MAIN();Is this expected behavior? Thanks! --Joana