我认为我已经找到了罪魁祸首:MKL(截至今天的最新版本)。我使用Pardiso,以下示例会缓慢泄漏:每13秒约0.1 MB,导致隔夜280 MB。这些是我从模拟中得到的数字。
如果您想尝试,请使用以下方式进行编译:
icpc -std=c++11 pardiso-leak.cpp -o main -lmkl_intel_lp64 -lmkl_core -lmkl_intel_thread -liomp5 -ldl -lpthread -lm
感谢大家的帮助。我已向英特尔报告了这个错误。
#include <iostream>
#include <vector>
#include "mkl_pardiso.h"
#include "mkl_types.h"
int main (int argc, char const *argv[])
{
const auto n = std::size_t{1000};
auto m = MKL_INT{n * n};
auto values = std::vector<double>();
auto column = std::vector<MKL_INT>();
auto row = std::vector<MKL_INT>();
row.push_back(1);
for(std::size_t j = 0; j < n; ++j) {
column.push_back(j + 1);
values.push_back(1.0);
column.push_back(j + n + 1);
values.push_back(0.1);
row.push_back(column.size() + 1);
}
for(std::size_t i = 1; i < n - 1; ++i) {
for(std::size_t j = 0; j < n; ++j) {
column.push_back(n * i + j - n + 1);
values.push_back(0.1);
column.push_back(n * i + j + 1);
values.push_back(1.0);
column.push_back(n * i + j + n + 1);
values.push_back(0.1);
row.push_back(column.size() + 1);
}
}
for(std::size_t j = 0; j < n; ++j) {
column.push_back((n - 1) * n + j - n + 1);
values.push_back(0.1);
column.push_back((n - 1) * n + j + 1);
values.push_back(1.0);
row.push_back(column.size() + 1);
}
auto y = std::vector<double>(m, 1.0);
auto x = std::vector<double>(m, 0.0);
auto pardiso_nrhs = MKL_INT{1};
auto pardiso_max_fact = MKL_INT{1};
auto pardiso_mnum = MKL_INT{1};
auto pardiso_mtype = MKL_INT{11};
auto pardiso_msglvl = MKL_INT{0};
MKL_INT pardiso_iparm[64];
for (int i = 0; i < 64; ++i) {
pardiso_iparm[i] = 0;
}
pardiso_iparm[0] = 1;
pardiso_iparm[1] = 2;
pardiso_iparm[3] = 0;
pardiso_iparm[4] = 0;
pardiso_iparm[5] = 0;
pardiso_iparm[7] = 0;
pardiso_iparm[8] = 0;
pardiso_iparm[9] = 13;
pardiso_iparm[10] = 1;
pardiso_iparm[11] = 0;
pardiso_iparm[12] = 1;
pardiso_iparm[17] = -1;
pardiso_iparm[18] = 0;
pardiso_iparm[20] = 0;
pardiso_iparm[23] = 1;
pardiso_iparm[24] = 0;
pardiso_iparm[26] = 0;
pardiso_iparm[27] = 0;
pardiso_iparm[30] = 0;
pardiso_iparm[31] = 0;
pardiso_iparm[32] = 0;
pardiso_iparm[33] = 0;
pardiso_iparm[34] = 0;
pardiso_iparm[59] = 0;
pardiso_iparm[60] = 0;
pardiso_iparm[61] = 0;
pardiso_iparm[62] = 0;
pardiso_iparm[63] = 0;
void* pardiso_pt[64];
for (int i = 0; i < 64; ++i) {
pardiso_pt[i] = nullptr;
}
auto error = MKL_INT{0};
auto phase = MKL_INT{11};
MKL_INT i_dummy;
double d_dummy;
PARDISO(pardiso_pt, &pardiso_max_fact, &pardiso_mnum, &pardiso_mtype,
&phase, &m, values.data(), row.data(), column.data(), &i_dummy,
&pardiso_nrhs, pardiso_iparm, &pardiso_msglvl, &d_dummy,
&d_dummy, &error);
phase = 22;
PARDISO(pardiso_pt, &pardiso_max_fact, &pardiso_mnum, &pardiso_mtype,
&phase, &m, values.data(), row.data(), column.data(), &i_dummy,
&pardiso_nrhs, pardiso_iparm, &pardiso_msglvl, &d_dummy,
&d_dummy, &error);
phase = 33;
for(size_t i = 0; i < 10000; ++i) {
std::cout << "i = " << i << std::endl;
PARDISO(pardiso_pt, &pardiso_max_fact, &pardiso_mnum, &pardiso_mtype,
&phase, &m, values.data(), row.data(), column.data(), &i_dummy,
&pardiso_nrhs, pardiso_iparm, &pardiso_msglvl, y.data(),
x.data(), &error);
}
phase = -1;
PARDISO(pardiso_pt, &pardiso_max_fact, &pardiso_mnum, &pardiso_mtype,
&phase, &m, values.data(), row.data(), column.data(), &i_dummy,
&pardiso_nrhs, pardiso_iparm, &pardiso_msglvl, &d_dummy,
&d_dummy, &error);
return 0;
}
std::vector
也是如此。我有一个调试模式,将所有的il::Vector<double>
初始化为 NaN。通过 包装std::vector
来实现这一点,而不是完全替换它。 - Lightness Races in Orbit