My Machine
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 96
On-line CPU(s) list: 0-95
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Gold 6252N CPU @ 2.30GHz
Stepping: 7
CPU MHz: 1699.871
CPU max MHz: 3600.0000
CPU min MHz: 1000.0000
BogoMIPS: 4600.00
Virtualization: VT-x
L1d cache: 32K
L1i cache: 32K
L2 cache: 1024K
L3 cache: 36608K
NUMA node0 CPU(s): 0-23,48-71
NUMA node1 CPU(s): 24-47,72-95
Core topology: two sockets, 24 cores per socket, 48 cores total
SMT status: enabled, but not utilized
Max clock rate: 1.7GHz (single-core and multicore)
Peak performance:
--single-core: 54.4 GFLOPS(double-precision)
--multicore: 54.4 GFLOPS/core (double-precision) 2611.2 GFLOPS/48 cores(double-precision)
I have fixed the frequency of the CPU at 1.7GHz by commands: sudo cpupower -c all frequency-set -u 1.7GHz, sudo cpupower -c all frequency-set -d 1.7GHz.
Code sample
int main(int argc, const char *argv[])
{
// matrix parameters: A * X = B, column major
int N, NRHS, LDA, LDB;
// test parameters
int N_START = 1000, N_END = 30000, NRHS_START = 1000, NRHS_END = 1000, INC = 1000, REPEAT = 3;
// N, NRHS, REPEAT=3
N = N_START, LDA = N, NRHS = NRHS_START, LDB = N;
double gflops[50][50][10];
while (N <= N_END){
while(NRHS <= NRHS_END){
double *A = NULL, *B = NULL;
int *IPIV = NULL;
for(int re_count = 0; re_count < REPEAT; ++ re_count)
{
A = (double *) malloc (sizeof(double) * N * N);
B = (double *) malloc (sizeof(double) * N * NRHS);
int seed[] = {0, 0, 0, 1};
LAPACKE_dlarnv(1, seed, N * N, A);
LAPACKE_dlarnv(1, seed, N * NRHS, B);
IPIV = (int *) malloc (sizeof(int) * N);
struct timeval start, finish;
gettimeofday(&start, NULL);
int info = LAPACKE_dgesv(LAPACK_COL_MAJOR, N, NRHS, A, LDA, IPIV, B, LDB);
gettimeofday(&finish, NULL);
if(info == 0){
double d_n = N, d_nrhs = NRHS;
double ops = ((2.0*d_n*d_n*d_n/3.0 - d_n*d_n/2.0 + 5.0*d_n/6.0) + (d_nrhs * (2*d_n*d_n - d_n))) * 1.0e-9;
gflops[N/INC - 1][NRHS/INC - 1][re_count] = ops / ( (finish.tv_sec - start.tv_sec) * 1.0 + (finish.tv_usec - start.tv_usec) * 1.0e-6 );
}
else{
fprintf(stderr, "[ERROR]: LAPACKE_dgesv failed\n");
exit(EXIT_FAILURE);
}
free(A), free(B), free(IPIV);
A = NULL, B = NULL, IPIV = NULL;
}
NRHS += INC;
}
N += INC, LDA = N, NRHS = NRHS_START, LDB = N;
sleep(10);
}
return 0;
}
[xx@cn0 code]$ export OMP_NUM_THREADS=48 GOMP_CPU_AFFINITY="0-47:1"
[xx@cn0 code]$ make test_dgesv_mkl.x
gcc -O2 -fopenmp -fPIC -o test_dgesv.o -c test_dgesv.c
gcc test_dgesv.o -L/home/xx/lib/intel/oneapi/mkl/2022.1.0/lib/intel64/ -lmkl_intel_lp64 -lmkl_core -lmkl_gnu_thread -lpthread -lm -ldl -fopenmp -o test_dgesv_mkl.x -lm -fopenmp -fPIC
[xx@cn0 code]$ numactl --interleave=all ./test_dgesv_mkl.x 
In the best case(N=27 000,NRHS=1 000), MKL can reach 65.64%(1714.609/2611.2) of the theoretical peak. Have I gotten the right results? Where can I find some relevant experimental results?
Regards,
lianchen.
链接已复制
