Grid IO benchmark cleanup
This commit is contained in:
parent
ce890a8fc2
commit
51eae5723e
@ -32,7 +32,7 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
|
||||
#ifdef HAVE_LIME
|
||||
using namespace Grid;
|
||||
|
||||
std::string filestem(const int l) { return "iobench_l" + std::to_string(l); }
|
||||
std::string filestem(const int l) { return "io/iobench_l" + std::to_string(l); }
|
||||
|
||||
int vol(const int i) { return BENCH_IO_LMIN + 2 * i; }
|
||||
|
||||
@ -56,13 +56,6 @@ template <typename Mat> void stats(Mat &mean, Mat &stdDev, const std::vector<Mat
|
||||
mean /= n;
|
||||
}
|
||||
|
||||
#define grid_printf(...) \
|
||||
{ \
|
||||
char _buf[1024]; \
|
||||
sprintf(_buf, __VA_ARGS__); \
|
||||
MSG << _buf; \
|
||||
}
|
||||
|
||||
enum
|
||||
{
|
||||
sRead = 0,
|
||||
@ -83,58 +76,58 @@ int main(int argc, char **argv)
|
||||
std::vector<Eigen::VectorXd> avPerf(BENCH_IO_NPASS, Eigen::VectorXd::Zero(4));
|
||||
std::vector<int> latt;
|
||||
|
||||
MSG << "Grid is setup to use " << threads << " threads" << std::endl;
|
||||
MSG << "MPI partition " << mpi << std::endl;
|
||||
GRID_MSG << "Grid is setup to use " << threads << " threads" << std::endl;
|
||||
GRID_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 << "Benchmark std write" << std::endl;
|
||||
MSG << SEP << std::endl;
|
||||
grid_big_sep();
|
||||
GRID_MSG << "Pass " << i + 1 << "/" << BENCH_IO_NPASS << std::endl;
|
||||
grid_big_sep();
|
||||
grid_small_sep();
|
||||
GRID_MSG << "Benchmark std write" << std::endl;
|
||||
grid_small_sep();
|
||||
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]};
|
||||
|
||||
MSG << "-- Local volume " << l << "^4" << std::endl;
|
||||
GRID_MSG << "-- Local volume " << l << "^4" << std::endl;
|
||||
writeBenchmark<LatticeFermion>(latt, filestem(l), stdWrite<LatticeFermion>);
|
||||
perf[i](volInd(l), sWrite) = BinaryIO::lastPerf.mbytesPerSecond;
|
||||
}
|
||||
|
||||
MSG << SEP << std::endl;
|
||||
MSG << "Benchmark std read" << std::endl;
|
||||
MSG << SEP << std::endl;
|
||||
grid_small_sep();
|
||||
GRID_MSG << "Benchmark std read" << std::endl;
|
||||
grid_small_sep();
|
||||
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]};
|
||||
|
||||
MSG << "-- Local volume " << l << "^4" << std::endl;
|
||||
GRID_MSG << "-- Local volume " << l << "^4" << std::endl;
|
||||
readBenchmark<LatticeFermion>(latt, filestem(l), stdRead<LatticeFermion>);
|
||||
perf[i](volInd(l), sRead) = BinaryIO::lastPerf.mbytesPerSecond;
|
||||
}
|
||||
|
||||
#ifdef HAVE_LIME
|
||||
MSG << SEP << std::endl;
|
||||
MSG << "Benchmark Grid C-Lime write" << std::endl;
|
||||
MSG << SEP << std::endl;
|
||||
grid_small_sep();
|
||||
GRID_MSG << "Benchmark Grid C-Lime write" << std::endl;
|
||||
grid_small_sep();
|
||||
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]};
|
||||
|
||||
MSG << "-- Local volume " << l << "^4" << std::endl;
|
||||
GRID_MSG << "-- Local volume " << l << "^4" << std::endl;
|
||||
writeBenchmark<LatticeFermion>(latt, filestem(l), limeWrite<LatticeFermion>);
|
||||
perf[i](volInd(l), gWrite) = BinaryIO::lastPerf.mbytesPerSecond;
|
||||
}
|
||||
|
||||
MSG << SEP << std::endl;
|
||||
MSG << "Benchmark Grid C-Lime read" << std::endl;
|
||||
MSG << SEP << std::endl;
|
||||
grid_small_sep();
|
||||
GRID_MSG << "Benchmark Grid C-Lime read" << std::endl;
|
||||
grid_small_sep();
|
||||
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]};
|
||||
|
||||
MSG << "-- Local volume " << l << "^4" << std::endl;
|
||||
GRID_MSG << "-- Local volume " << l << "^4" << std::endl;
|
||||
readBenchmark<LatticeFermion>(latt, filestem(l), limeRead<LatticeFermion>);
|
||||
perf[i](volInd(l), gRead) = BinaryIO::lastPerf.mbytesPerSecond;
|
||||
}
|
||||
@ -159,13 +152,13 @@ int main(int argc, char **argv)
|
||||
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_big_sep();
|
||||
GRID_MSG << "SUMMARY" << std::endl;
|
||||
grid_big_sep();
|
||||
GRID_MSG << "Summary of individual results (all results in MB/s)." << std::endl;
|
||||
GRID_MSG << "Every second colum gives the standard deviation of the previous column."
|
||||
<< std::endl;
|
||||
GRID_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)
|
||||
@ -176,10 +169,10 @@ int main(int argc, char **argv)
|
||||
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_MSG << std::endl;
|
||||
GRID_MSG << "Robustness of individual results, in %. (rob = 100% - std dev / mean)"
|
||||
<< std::endl;
|
||||
GRID_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)
|
||||
@ -187,21 +180,21 @@ int main(int argc, char **argv)
|
||||
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_MSG << std::endl;
|
||||
GRID_MSG << "Summary of results averaged over local volumes 24^4-" << BENCH_IO_LMAX
|
||||
<< "^4 (all results in MB/s)." << std::endl;
|
||||
GRID_MSG << "Every second colum gives the standard deviation of the previous column."
|
||||
<< std::endl;
|
||||
GRID_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_MSG << std::endl;
|
||||
GRID_MSG << "Robustness of volume-averaged results, in %. (rob = 100% - std dev / mean)"
|
||||
<< std::endl;
|
||||
GRID_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),
|
||||
|
||||
@ -18,12 +18,8 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
|
||||
#ifndef Benchmark_IO_hpp_
|
||||
#define Benchmark_IO_hpp_
|
||||
|
||||
#include "Common.hpp"
|
||||
#include <Grid/Grid.h>
|
||||
#define MSG std::cout << GridLogMessage
|
||||
#define SEP \
|
||||
"-----------------------------------------------------------------------------"
|
||||
#define BIGSEP \
|
||||
"============================================================================="
|
||||
#ifdef HAVE_LIME
|
||||
|
||||
namespace Grid
|
||||
@ -50,9 +46,9 @@ namespace Grid
|
||||
// crc = GridChecksum::crc32(vec_v.cpu_ptr, size);
|
||||
// std::fwrite(&crc, sizeof(uint32_t), 1, file);
|
||||
// crcWatch.Stop();
|
||||
// MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
|
||||
// ioWatch.Start();
|
||||
// std::fwrite(vec_v.cpu_ptr, sizeof(typename Field::scalar_object),
|
||||
// GRID_MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec <<
|
||||
// std::endl; ioWatch.Start(); std::fwrite(vec_v.cpu_ptr, sizeof(typename
|
||||
// Field::scalar_object),
|
||||
// vec.Grid()->lSites(), file);
|
||||
// ioWatch.Stop();
|
||||
// std::fclose(file);
|
||||
@ -61,11 +57,11 @@ namespace Grid
|
||||
// p.size = size;
|
||||
// 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()
|
||||
// GRID_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;
|
||||
// GRID_MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
|
||||
// }
|
||||
|
||||
// template <typename Field> void stdRead(Field &vec, const std::string filestem)
|
||||
@ -94,16 +90,14 @@ namespace Grid
|
||||
// crcData = GridChecksum::crc32(vec_v.cpu_ptr, size);
|
||||
// crcWatch.Stop();
|
||||
// }
|
||||
// MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
|
||||
// assert(crcData == crcRead);
|
||||
// size *= vec.Grid()->ProcessorCount();
|
||||
// auto &p = BinaryIO::lastPerf;
|
||||
// p.size = size;
|
||||
// p.time = ioWatch.useconds();
|
||||
// GRID_MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec <<
|
||||
// std::endl; assert(crcData == crcRead); size *= vec.Grid()->ProcessorCount(); auto
|
||||
// &p = BinaryIO::lastPerf; p.size = size; 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() << ", "
|
||||
// GRID_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;
|
||||
// GRID_MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
|
||||
// }
|
||||
|
||||
template <typename Field> void stdWrite(const std::string filestem, Field &vec)
|
||||
@ -122,7 +116,7 @@ namespace Grid
|
||||
crc = GridChecksum::crc32(vec_v.cpu_ptr, size);
|
||||
file.write(reinterpret_cast<char *>(&crc), sizeof(uint32_t) / sizeof(char));
|
||||
crcWatch.Stop();
|
||||
MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
|
||||
GRID_MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
|
||||
ioWatch.Start();
|
||||
file.write(reinterpret_cast<char *>(vec_v.cpu_ptr), sizec);
|
||||
file.flush();
|
||||
@ -132,9 +126,9 @@ namespace Grid
|
||||
p.size = size;
|
||||
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;
|
||||
GRID_MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed()
|
||||
<< ", " << p.mbytesPerSecond << " MB/s" << std::endl;
|
||||
GRID_MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
|
||||
}
|
||||
|
||||
template <typename Field> void stdRead(Field &vec, const std::string filestem)
|
||||
@ -163,16 +157,17 @@ namespace Grid
|
||||
crcData = GridChecksum::crc32(vec_v.cpu_ptr, size);
|
||||
crcWatch.Stop();
|
||||
}
|
||||
MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
|
||||
GRID_MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec
|
||||
<< std::endl;
|
||||
assert(crcData == crcRead);
|
||||
size *= vec.Grid()->ProcessorCount();
|
||||
auto &p = BinaryIO::lastPerf;
|
||||
p.size = size;
|
||||
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;
|
||||
GRID_MSG << "Std I/O read: Read " << p.size << " bytes in " << ioWatch.Elapsed()
|
||||
<< ", " << p.mbytesPerSecond << " MB/s" << std::endl;
|
||||
GRID_MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
|
||||
}
|
||||
|
||||
template <typename Field> void limeWrite(const std::string filestem, Field &vec)
|
||||
|
||||
Loading…
Reference in New Issue
Block a user