portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

multi-pass I/O benchmark, with statistic and robustness summary

Browse Source
This commit is contained in:
Antonin Portelli 2020-10-09 20:29:40 +01:00
parent b24a504d7c
commit 22caa158cc
2 changed files with 198 additions and 58 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

134
benchmarks/Benchmark_IO.cc
View File

@ -1,8 +1,16 @@
#include "Benchmark_IO.hpp" #include "Benchmark_IO.hpp"
#ifndef BENCH_IO_LMIN
#define BENCH_IO_LMIN 20
#endif
#ifndef BENCH_IO_LMAX #ifndef BENCH_IO_LMAX
#define BENCH_IO_LMAX 40 #define BENCH_IO_LMAX 30
#endif
#ifndef BENCH_IO_NPASS
#define BENCH_IO_NPASS 10
#endif #endif
using namespace Grid; using namespace Grid;
@ -12,62 +20,174 @@ std::string filestem(const int l)
return "iobench_l" + std::to_string(l); return "iobench_l" + std::to_string(l);
} }
int vol(const int i)
{
return BENCH_IO_LMIN + 2*i;
}
int volInd(const int l)
{
return (l - BENCH_IO_LMIN)/2;
}
template <typename Mat>
void stats(Mat &mean, Mat &stdDev, const std::vector<Mat> &data)
{
auto nr = data[0].rows(), nc = data[0].cols();
Eigen::MatrixXd sqSum(nr, nc);
double n = static_cast<double>(data.size());
assert(n > 1.);
mean = Mat::Zero(nr, nc);
sqSum = Mat::Zero(nr, nc);
for (auto &d: data)
{
mean += d;
sqSum += d.cwiseProduct(d);
}
stdDev = ((sqSum - mean.cwiseProduct(mean)/n)/(n - 1.)).cwiseSqrt();
mean /= n;
}
#define grid_printf(...) \
MSG << "";\
printf(__VA_ARGS__);
enum {sRead = 0, sWrite = 1, gRead = 2, gWrite = 3};
int main (int argc, char ** argv) int main (int argc, char ** argv)
{ {
Grid_init(&argc,&argv); Grid_init(&argc,&argv);
int64_t threads = GridThread::GetThreads(); int64_t threads = GridThread::GetThreads();
auto mpi = GridDefaultMpi(); auto mpi = GridDefaultMpi();
unsigned int nVol = (BENCH_IO_LMAX - BENCH_IO_LMIN)/2 + 1;
unsigned int nRelVol = (BENCH_IO_LMAX - 24)/2 + 1;
std::vector<Eigen::MatrixXd> perf(BENCH_IO_NPASS, Eigen::MatrixXd::Zero(nVol, 4));
std::vector<Eigen::VectorXd> avPerf(BENCH_IO_NPASS, Eigen::VectorXd::Zero(4));
std::vector<int> latt; std::vector<int> latt;
MSG << "Grid is setup to use " << threads << " threads" << std::endl; MSG << "Grid is setup to use " << threads << " threads" << std::endl;
MSG << "MPI partition " << mpi << std::endl; MSG << "MPI partition " << mpi << std::endl;
for (unsigned int i = 0; i < BENCH_IO_NPASS; ++i)
{
MSG << BIGSEP << std::endl;
MSG << "Pass " << i + 1 << "/" << BENCH_IO_NPASS << std::endl;
MSG << BIGSEP << std::endl;
MSG << SEP << std::endl; MSG << SEP << std::endl;
MSG << "Benchmark std write" << std::endl; MSG << "Benchmark std write" << std::endl;
MSG << SEP << std::endl; MSG << SEP << std::endl;
for (int l = 4; l <= BENCH_IO_LMAX; l += 2) for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{ {
latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]}; latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
MSG << "-- Local volume " << l << "^4" << std::endl; MSG << "-- Local volume " << l << "^4" << std::endl;
writeBenchmark<LatticeFermion>(latt, filestem(l), stdWrite<LatticeFermion>); writeBenchmark<LatticeFermion>(latt, filestem(l), stdWrite<LatticeFermion>);
perf[i](volInd(l), sWrite) = BinaryIO::lastPerf.mbytesPerSecond;
} }
MSG << SEP << std::endl; MSG << SEP << std::endl;
MSG << "Benchmark std read" << std::endl; MSG << "Benchmark std read" << std::endl;
MSG << SEP << std::endl; MSG << SEP << std::endl;
for (int l = 4; l <= BENCH_IO_LMAX; l += 2) for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{ {
latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]}; latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
MSG << "-- Local volume " << l << "^4" << std::endl; MSG << "-- Local volume " << l << "^4" << std::endl;
readBenchmark<LatticeFermion>(latt, filestem(l), stdRead<LatticeFermion>); readBenchmark<LatticeFermion>(latt, filestem(l), stdRead<LatticeFermion>);
perf[i](volInd(l), sRead) = BinaryIO::lastPerf.mbytesPerSecond;
} }
#ifdef HAVE_LIME #ifdef HAVE_LIME
MSG << SEP << std::endl; MSG << SEP << std::endl;
MSG << "Benchmark Grid C-Lime write" << std::endl; MSG << "Benchmark Grid C-Lime write" << std::endl;
MSG << SEP << std::endl; MSG << SEP << std::endl;
for (int l = 4; l <= BENCH_IO_LMAX; l += 2) for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{ {
latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]}; latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
MSG << "-- Local volume " << l << "^4" << std::endl; MSG << "-- Local volume " << l << "^4" << std::endl;
writeBenchmark<LatticeFermion>(latt, filestem(l), limeWrite<LatticeFermion>); writeBenchmark<LatticeFermion>(latt, filestem(l), limeWrite<LatticeFermion>);
perf[i](volInd(l), gWrite) = BinaryIO::lastPerf.mbytesPerSecond;
} }
MSG << SEP << std::endl; MSG << SEP << std::endl;
MSG << "Benchmark Grid C-Lime read" << std::endl; MSG << "Benchmark Grid C-Lime read" << std::endl;
MSG << SEP << std::endl; MSG << SEP << std::endl;
for (int l = 4; l <= BENCH_IO_LMAX; l += 2) for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{ {
latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]}; latt = {l*mpi[0], l*mpi[1], l*mpi[2], l*mpi[3]};
MSG << "-- Local volume " << l << "^4" << std::endl; MSG << "-- Local volume " << l << "^4" << std::endl;
readBenchmark<LatticeFermion>(latt, filestem(l), limeRead<LatticeFermion>); readBenchmark<LatticeFermion>(latt, filestem(l), limeRead<LatticeFermion>);
perf[i](volInd(l), gRead) = BinaryIO::lastPerf.mbytesPerSecond;
} }
#endif #endif
avPerf[i].fill(0.);
for (int f = 0; f < 4; ++f)
for (int l = 24; l <= BENCH_IO_LMAX; l += 2)
{
avPerf[i](f) += perf[i](volInd(l), f);
}
avPerf[i] /= nRelVol;
}
Eigen::MatrixXd mean(nVol, 4), stdDev(nVol, 4), rob(nVol, 4);
Eigen::VectorXd avMean(4), avStdDev(4), avRob(4);
double n = BENCH_IO_NPASS;
stats(mean, stdDev, perf);
stats(avMean, avStdDev, avPerf);
rob.fill(100.);
rob -= 100.*stdDev.cwiseQuotient(mean.cwiseAbs());
avRob.fill(100.);
avRob -= 100.*avStdDev.cwiseQuotient(avMean.cwiseAbs());
MSG << BIGSEP << std::endl;
MSG << "SUMMARY" << std::endl;
MSG << BIGSEP << std::endl;
MSG << "Summary of individual results (all results in MB/s)." << std::endl;
MSG << "Every second colum gives the standard deviation of the previous column." << std::endl;
MSG << std::endl;
grid_printf("%4s %12s %12s %12s %12s %12s %12s %12s %12s\n",
"L", "std read", "std dev", "std write", "std dev",
"Grid read", "std dev", "Grid write", "std dev");
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{
grid_printf("%4d %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f\n",
l, mean(volInd(l), sRead), stdDev(volInd(l), sRead),
mean(volInd(l), sWrite), stdDev(volInd(l), sWrite),
mean(volInd(l), gRead), stdDev(volInd(l), gRead),
mean(volInd(l), gWrite), stdDev(volInd(l), gWrite));
}
MSG << std::endl;
MSG << "Robustness of individual results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
MSG << std::endl;
grid_printf("%4s %12s %12s %12s %12s\n",
"L", "std read", "std write", "Grid read", "Grid write");
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{
grid_printf("%4d %12.1f %12.1f %12.1f %12.1f\n",
l, rob(volInd(l), sRead), rob(volInd(l), sWrite),
rob(volInd(l), gRead), rob(volInd(l), gWrite));
}
MSG << std::endl;
MSG << "Summary of results averaged over local volumes 24^4-" << BENCH_IO_LMAX << "^4 (all results in MB/s)." << std::endl;
MSG << "Every second colum gives the standard deviation of the previous column." << std::endl;
MSG << std::endl;
grid_printf("%12s %12s %12s %12s %12s %12s %12s %12s\n",
"std read", "std dev", "std write", "std dev",
"Grid read", "std dev", "Grid write", "std dev");
grid_printf("%12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f\n",
avMean(sRead), avStdDev(sRead), avMean(sWrite), avStdDev(sWrite),
avMean(gRead), avStdDev(gRead), avMean(gWrite), avStdDev(gWrite));
MSG << std::endl;
MSG << "Robustness of volume-averaged results, in \%. (rob = 100\% - std dev / mean)" << std::endl;
MSG << std::endl;
grid_printf("%12s %12s %12s %12s\n",
"std read", "std write", "Grid read", "Grid write");
grid_printf("%12.1f %12.1f %12.1f %12.1f\n",
avRob(sRead), avRob(sWrite), avRob(gRead), avRob(gWrite));
Grid_finalize(); Grid_finalize();

44
benchmarks/Benchmark_IO.hpp
View File

@ -5,6 +5,8 @@
#ifdef HAVE_LIME #ifdef HAVE_LIME
#define MSG std::cout << GridLogMessage #define MSG std::cout << GridLogMessage
#define SEP \ #define SEP \
"-----------------------------------------------------------------------------"
#define BIGSEP \
"=============================================================================" "============================================================================="
namespace Grid { namespace Grid {
@ -37,9 +39,12 @@ using ReaderFn = std::function<void(Field &, const std::string)>;
// ioWatch.Stop(); // ioWatch.Stop();
// std::fclose(file); // std::fclose(file);
// size *= vec.Grid()->ProcessorCount(); // size *= vec.Grid()->ProcessorCount();
// MSG << "Std I/O write: Wrote " << size << " bytes in " << ioWatch.Elapsed() // auto &p = BinaryIO::lastPerf;
// << ", performance " << size/1024./1024./(ioWatch.useconds()/1.e6) // p.size = size;
// << " MB/s" << std::endl; // p.time = ioWatch.useconds();
// p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
// MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed()
// << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
// MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl; // MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
// } // }
// //
@ -72,9 +77,12 @@ using ReaderFn = std::function<void(Field &, const std::string)>;
// MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl; // MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
// assert(crcData == crcRead); // assert(crcData == crcRead);
// size *= vec.Grid()->ProcessorCount(); // size *= vec.Grid()->ProcessorCount();
// MSG << "Std I/O read: Read " << size << " bytes in " << ioWatch.Elapsed() // auto &p = BinaryIO::lastPerf;
// << ", performance " << size/1024./1024./(ioWatch.useconds()/1.e6) // p.size = size;
// << " MB/s" << std::endl; // p.time = ioWatch.useconds();
// p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
// MSG << "Std I/O read: Read " << p.size << " bytes in " << ioWatch.Elapsed()
// << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
// MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl; // MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
// } // }
@ -100,9 +108,12 @@ void stdWrite(const std::string filestem, Field &vec)
file.flush(); file.flush();
ioWatch.Stop(); ioWatch.Stop();
size *= vec.Grid()->ProcessorCount(); size *= vec.Grid()->ProcessorCount();
MSG << "Std I/O write: Wrote " << size << " bytes in " << ioWatch.Elapsed() auto &p = BinaryIO::lastPerf;
<< ", " << size/1024./1024./(ioWatch.useconds()/1.e6) p.size = size;
<< " MB/s" << std::endl; p.time = ioWatch.useconds();
p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed()
<< ", " << p.mbytesPerSecond << " MB/s" << std::endl;
MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl; MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
} }
@ -135,9 +146,12 @@ void stdRead(Field &vec, const std::string filestem)
MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl; MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
assert(crcData == crcRead); assert(crcData == crcRead);
size *= vec.Grid()->ProcessorCount(); size *= vec.Grid()->ProcessorCount();
MSG << "Std I/O read: Read " << size << " bytes in " << ioWatch.Elapsed() auto &p = BinaryIO::lastPerf;
<< ", " << size/1024./1024./(ioWatch.useconds()/1.e6) p.size = size;
<< " MB/s" << std::endl; p.time = ioWatch.useconds();
p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
MSG << "Std I/O read: Read " << p.size << " bytes in " << ioWatch.Elapsed()
<< ", " << p.mbytesPerSecond << " MB/s" << std::endl;
MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl; MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
} }
@ -200,6 +214,7 @@ void writeBenchmark(const Coordinate &latt, const std::string filename,
auto simd = GridDefaultSimd(latt.size(), Field::vector_type::Nsimd()); auto simd = GridDefaultSimd(latt.size(), Field::vector_type::Nsimd());
std::shared_ptr<GridCartesian> gBasePt(SpaceTimeGrid::makeFourDimGrid(latt, simd, mpi)); std::shared_ptr<GridCartesian> gBasePt(SpaceTimeGrid::makeFourDimGrid(latt, simd, mpi));
std::shared_ptr<GridBase> gPt; std::shared_ptr<GridBase> gPt;
std::random_device rd;
makeGrid(gPt, gBasePt, Ls, rb); makeGrid(gPt, gBasePt, Ls, rb);
@ -207,6 +222,11 @@ void writeBenchmark(const Coordinate &latt, const std::string filename,
GridParallelRNG rng(g); GridParallelRNG rng(g);
Field vec(g); Field vec(g);
rng.SeedFixedIntegers({static_cast<int>(rd()), static_cast<int>(rd()),
static_cast<int>(rd()), static_cast<int>(rd()),
static_cast<int>(rd()), static_cast<int>(rd()),
static_cast<int>(rd()), static_cast<int>(rd())});
random(rng, vec); random(rng, vec);
write(filename, vec); write(filename, vec);
} }