'Hybrid MPI+OpenMP Vs MPI Performance
I am converting a 3-D Jacobi solver from pure MPI to Hybrid MPI+OpenMP. I have a 192x192x192 array which is divided among 24 processes in Pure MPI in 1-D decomposition i.e. each process has 192/24 x 192 x 192 = 8 x 192 x 192 slab of data. Now I do :
for(i=0 ; i <= 7; i++)
for(j=0; j<= 191; j++)
for(k=0; k<= 191; k++)
{
unew[i][j][k] = 1/6.0 * (u[i+1][j][k]+u[i-1][j][k]+
u[i][j+1][k]+u[i][j-1][k]+
u[i][j][k+1]+u[i][j][k-1]);
}
This update takes around 60 seconds for each process.
Now with Hybrid MPI, I run two processes (1 process per socket --bind-to socket --map-by socket and OMP_PROC_PLACES=coreswith OMP_PROC_BIND=close). I create 12 threads per MPI Process (i.e. 12 threads per socket or processor). Now each MPI process has an array of size : 192/2 x 192 x 192 = 96x192x192 elements. Each thread works on 96/12 x 192 x 192 = 8 x 192 x 192 portion of the array owned by each process. I do the same triple loop update using threads but the time is approximately 76 seconds for each thread. The load balance is perfect in both the problems. What could be the possible causes of performance degradation ? Is is False Sharing because threads could be invalidating the cache lines close to each other's chunk of data ? If yes, then how do I reduce this performance degradation ? (I have purposefully not mentioned ghost data but initially I am NOT overlapping communication with computation.)
In response to the comments below, am posting the code. Apologies for the long MWE but you can very safely ignore (1) Header files declaration (2) Variable Declaration (3) Memory allocation routine (4) Formation of Cartesian Topology (5) Setting boundary conditions in parallel using OpenMP parallel region (6) Declaration of MPI_Type_subarray datatype (7) MPI_Isend() and MPI_Irecv() calls and just concentrate on (a) INDEPENDENT UPDATE OpenMP parallel region (b) independent_update(...) routine being called from here.
/* IGNORE THIS PORTION */
#include<mpi.h>
#include<omp.h>
#include<stdio.h>
#include<stdlib.h>
#include<math.h>
#define MIN(a,b) (a < b ? a : b)
#define Tol 0.00001
/* IGNORE THIS ROUTINE */
void input(int *X, int *Y, int *Z)
{
int a=193, b=193, c=193;
*X = a;
*Y = b;
*Z = c;
}
/* IGNORE THIS ROUTINE */
float*** allocate_mem(int X, int Y, int Z)
{
int i,j;
float ***matrix;
float *arr;
arr = (float*)calloc(X*Y*Z, sizeof(float));
matrix = (float***)calloc(X, sizeof(float**));
for(i = 0 ; i<= X-1; i++)
matrix[i] = (float**)calloc(Y, sizeof(float*));
for(i = 0 ; i <= X-1; i++)
for(j=0; j<= Y-1; j++)
matrix[i][j] = &(arr[i*Y*Z + j*Z]);
return matrix ;
}
/* THIS ROUTINE IS IMPORTANT */
float independent_update(float ***old, float ***new, int NX, int NY, int NZ, int tID, int chunk)
{
int i,j,k, start, end;
float error = 0.0;
float diff;
start = tID * chunk + 1;
end = MIN( (tID+1)*chunk, NX-2 );
for(i = start; i <= end ; i++)
{
for(j = 1; j<= NY-2; j++)
{
#pragma omp simd
for(k = 1; k<= NZ-2; k++)
{
new[i][j][k] = (1/6.0) *(old[i-1][j][k] + old[i+1][j][k] + old[i][j-1][k] + old[i][j+1][k] + old[i][j][k-1] + old[i][j][k+1] );
diff = 1.0 - new[i][j][k];
diff = (diff > 0 ? diff : -1.0 * diff );
if(diff > error)
error = diff;
}
}
}
return error;
}
int main(int argc, char *argv[])
{
/* IGNORE VARIABLE DECLARATION */
int size, rank; //Size of old_comm and rank of process
int i, j, k,l; //General loop variables
MPI_Comm old_comm, new_comm; //MPI_COMM_WORLD handle and for MPI_Cart_create()
int N[3]; //For taking input of size of matrix from user
int P; //Represent number of processes i.e. same as size
int dims[3]; //For dimensions of Cartesian topology
int PX, PY, PZ; //X dim, Y dim, Z dim of each process
float ***old, ***new, ***temp; //Matrices for results dimensions is (Px+2)*(PY+2)*(PZ+2)
int period[3]; //Periodicity for each dimension
int reorder; //Whether processes should be reordered in new cartesian topology
int ndims; //Number of dimensions (which is 3)
int Z_TOWARDS_U, Z_AWAY_U; //Z neighbour towards you and away from you (Z const)
int X_DOWN, X_UP; //Below plane and above plane (X const)
int Y_LEFT, Y_RIGHT; //Left plane and right plane (Y const)
int coords[3]; //Finding coordinates of processes
int dimension; //Used in MPI_Cart_shift() , values = 0, 1,2
int displacement; //Used in MPI_Cart_shift(), values will be +1 to find immediate neighbours
float l_max_err; //Local maximum error on process
float l_max_err_new; //For dependent faces.
float G_max_err = 1.0; //Maximum error for stopping criterion
int iterations = 0 ; //Counting number of iterations
MPI_Request send[6], recv[6]; //For MPI_Isend and MPI_Irecv
int start[3]; //Start will be defined in MPI_Isend() and MPI_Irecv()
int gsize[3]; //Defining global size of subarray
MPI_Datatype x_subarray; //For sending X_UP and X_DOWN
int local_x[3]; //Defining local plane size for X_UP/X_DOWN
MPI_Datatype y_subarray; //For sending Y_LEFT and Y_RIGHT
int local_y[3]; //Defining local plane for Y_LEFT/Y_RIGHT
MPI_Datatype z_subarray; //For sending Z_TOWARDS_U and Z_AWAY_U
int local_z[3]; //Defining local plan size for XY plane i.e. where Z=0
double strt, end; //For measuring time
double strt1, end1, delta1; //For measuring trivial time 1
double strt2, end2, delta2; //For measuring trivial time 2
double t_i_strt, t_i_end, t_i_sum=0; //Time for independent computational kernel
double t_up_strt, t_up_end, t_up_sum=0; //Time for X_UP
double t_down_strt, t_down_end, t_down_sum=0; //Time for X_DOWN
double t_left_strt, t_left_end, t_left_sum=0; //Time for Y_LEFT
double t_right_strt, t_right_end, t_right_sum=0; //Time for Y_RIGHT
double t_towards_strt, t_towards_end, t_towards_sum=0; //For Z_TOWARDS_U
double t_away_strt, t_away_end, t_away_sum=0; //For Z_AWAY_U
double t_comm_strt, t_comm_end, t_comm_sum=0; //Time comm + independent update (need to subtract to get comm time)
double t_setup_strt,t_setup_end; //Set-up start and end time
double t_allred_strt,t_allred_end,t_allred_total=0.0; //Measuring Allreduce time separately.
int threadID; //ID of a thread
int nthreads; //Total threads in OpenMP region
int chunk; //chunk - used to calculate iterations of a thread
/* IGNORE MPI STARTUP ETC */
MPI_Init(&argc, &argv);
t_setup_strt = MPI_Wtime();
old_comm = MPI_COMM_WORLD;
MPI_Comm_size(old_comm, &size);
MPI_Comm_rank(old_comm, &rank);
P = size;
if(rank == 0)
{
input(&N[0], &N[1], &N[2]);
}
MPI_Bcast(N, 3, MPI_INT, 0, old_comm);
dims[0] = 0;
dims[1] = 0;
dims[2] = 0;
period[0] = period[1] = period[2] = 0; //All dimensions aperiodic
reorder = 0 ; //No reordering of ranks in new_comm
ndims = 3;
MPI_Dims_create(P,ndims,dims);
MPI_Cart_create(old_comm, ndims, dims, period, reorder, &new_comm);
if( (N[0]-1) % dims[0] == 0 && (N[1]-1) % dims[1] == 0 && (N[2]-1) % dims[2] == 0 )
{
PX = (N[0]-1)/dims[0]; //Rows of unknowns each process gets
PY = (N[1]-1)/dims[1]; //Columns of unknowns each process gets
PZ = (N[2]-1)/dims[2]; //Depth of unknowns each process gets
}
old = allocate_mem(PX+2, PY+2, PZ+2); //3D arrays with ghost points
new = allocate_mem(PX+2, PY+2, PZ+2); //3D arrays with ghost points
dimension = 0;
displacement = 1;
MPI_Cart_shift(new_comm, dimension, displacement, &X_UP, &X_DOWN); //Find UP and DOWN neighbours
dimension = 1;
MPI_Cart_shift(new_comm, dimension, displacement, &Y_LEFT, &Y_RIGHT); //Find UP and DOWN neighbours
dimension = 2;
MPI_Cart_shift(new_comm, dimension, displacement, &Z_TOWARDS_U, &Z_AWAY_U); //Find UP and DOWN neighbours
/* IGNORE BOUNDARY SETUPS FOR PDE */
#pragma omp parallel for default(none) shared(old,new,PX,PY,PZ) private(i,j,k) schedule(static)
for(i = 0; i <= PX+1; i++)
{
for(j = 0; j <= PY+1; j++)
{
for(k = 0; k <= PZ+1; k++)
{
old[i][j][k] = 0.0;
new[i][j][k] = 0.0;
}
}
}
#pragma omp parallel default(none) shared(X_DOWN,X_UP,Y_LEFT,Y_RIGHT,Z_TOWARDS_U,Z_AWAY_U,old,new,PX,PY,PZ) private(i,j,k,threadID,nthreads)
{
threadID = omp_get_thread_num();
nthreads = omp_get_num_threads();
if(threadID == 0)
{
if(X_DOWN == MPI_PROC_NULL) //X is constant here, this is YZ upper plane
{
for(j = 1 ; j<= PY ; j++)
for(k = 1 ; k<= PZ ; k++)
{
old[0][j][k] = 1;
new[0][j][k] = 1; //Set boundaries in new also
}
}
}
if(threadID == (nthreads-1))
{
if(X_UP == MPI_PROC_NULL) //YZ lower plane
{
for(j = 1 ; j<= PY ; j++)
for(k = 1; k<= PZ ; k++)
{
old[PX+1][j][k] = 1;
new[PX+1][j][k] = 1;
}
}
}
if(Y_LEFT == MPI_PROC_NULL) //Y is constant, this is left XZ plane, possibly can use collapse(2)
{
#pragma omp for schedule(static)
for(i = 1 ; i<= PX ; i++)
for(k = 1; k<= PZ; k++)
{
old[i][0][k] = 1;
new[i][0][k] = 1;
}
}
if(Y_RIGHT == MPI_PROC_NULL) //XZ right plane, again collapse(2) potential
{
#pragma omp for schedule(static)
for(i = 1 ; i<= PX; i++)
for(k = 1; k<= PZ ; k++)
{
old[i][PY+1][k] = 1;
new[i][PY+1][k] = 1;
}
}
if(Z_TOWARDS_U == MPI_PROC_NULL) //Z is constant here, towards you XY plane, collapse(2)
{
#pragma omp for schedule(static)
for(i = 1 ; i<= PX ; i++)
for(j = 1; j<= PY ; j++)
{
old[i][j][0] = 1;
new[i][j][0] = 1;
}
}
if(Z_AWAY_U == MPI_PROC_NULL) //Away from you XY plane, collapse(2)
{
#pragma omp for schedule(static)
for(i = 1 ; i<= PX; i++)
for(j = 1; j<= PY ; j++)
{
old[i][j][PZ+1] = 1;
new[i][j][PZ+1] = 1;
}
}
}
/* IGNORE SUBARRAY DECLARATION */
gsize[0] = PX+2; //Global sizes of 3-D cubes for each process
gsize[1] = PY+2;
gsize[2] = PZ+2;
start[0] = 0; //Will specify starting location while sending/receiving
start[1] = 0;
start[2] = 0;
local_x[0] = 1;
local_x[1] = PY;
local_x[2] = PZ;
MPI_Type_create_subarray(ndims, gsize, local_x, start, MPI_ORDER_C, MPI_FLOAT, &x_subarray);
MPI_Type_commit(&x_subarray);
local_y[0] = PX;
local_y[1] = 1;
local_y[2] = PZ;
MPI_Type_create_subarray(ndims, gsize, local_y, start, MPI_ORDER_C, MPI_FLOAT, &y_subarray);
MPI_Type_commit(&y_subarray);
local_z[0] = PX;
local_z[1] = PY;
local_z[2] = 1;
MPI_Type_create_subarray(ndims, gsize, local_z, start, MPI_ORDER_C, MPI_FLOAT, &z_subarray);
MPI_Type_commit(&z_subarray);
t_setup_end = MPI_Wtime();
strt = MPI_Wtime();
while(G_max_err > Tol) //iterations < ITERATIONS)
{
iterations++ ;
t_comm_strt = MPI_Wtime();
/* IGNORE MPI COMMUNICATION */
MPI_Irecv(&old[0][1][1], 1, x_subarray, X_DOWN, 10, new_comm, &recv[0]);
MPI_Irecv(&old[PX+1][1][1], 1, x_subarray, X_UP, 20, new_comm, &recv[1]);
MPI_Irecv(&old[1][PY+1][1], 1, y_subarray, Y_RIGHT, 30, new_comm, &recv[2]);
MPI_Irecv(&old[1][0][1], 1, y_subarray, Y_LEFT, 40, new_comm, &recv[3]);
MPI_Irecv(&old[1][1][PZ+1], 1, z_subarray, Z_AWAY_U, 50, new_comm, &recv[4]);
MPI_Irecv(&old[1][1][0], 1, z_subarray, Z_TOWARDS_U, 60, new_comm, &recv[5]);
MPI_Isend(&old[PX][1][1], 1, x_subarray, X_UP, 10, new_comm, &send[0]);
MPI_Isend(&old[1][1][1], 1, x_subarray, X_DOWN, 20, new_comm, &send[1]);
MPI_Isend(&old[1][1][1], 1, y_subarray, Y_LEFT, 30, new_comm, &send[2]);
MPI_Isend(&old[1][PY][1], 1, y_subarray, Y_RIGHT, 40, new_comm, &send[3]);
MPI_Isend(&old[1][1][1], 1, z_subarray, Z_TOWARDS_U, 50, new_comm, &send[4]);
MPI_Isend(&old[1][1][PZ], 1, z_subarray, Z_AWAY_U, 60, new_comm, &send[5]);
MPI_Waitall(6, send, MPI_STATUSES_IGNORE);
MPI_Waitall(6, recv, MPI_STATUSES_IGNORE);
t_comm_end = MPI_Wtime();
t_comm_sum = t_comm_sum + (t_comm_end - t_comm_strt);
/* Use threads in Independent update */
t_i_strt = MPI_Wtime();
l_max_err = 0.0; //Very important, Reduction result is combined with this !
/* THIS IS THE IMPORTANT REGION */
#pragma omp parallel default(none) shared(old,new,PX,PY,PZ,chunk) private(threadID,nthreads) reduction(max:l_max_err)
{
nthreads = omp_get_num_threads();
threadID = omp_get_thread_num();
chunk = (PX-1+1) / nthreads ;
l_max_err = independent_update(old, new, PX+2, PY+2, PZ+2, threadID, chunk);
}
t_i_end = MPI_Wtime();
t_i_sum = t_i_sum + (t_i_end - t_i_strt) ;
/* IGNORE THE REMAINING CODE */
t_allred_strt = MPI_Wtime();
MPI_Allreduce(&l_max_err, &G_max_err, 1, MPI_FLOAT, MPI_MAX, new_comm);
t_allred_end = MPI_Wtime();
t_allred_total = t_allred_total + (t_allred_end - t_allred_strt);
temp = new ;
new = old;
old = temp;
}
MPI_Barrier(new_comm);
end = MPI_Wtime();
if( rank == 0)
{
printf("\nIterations = %d, G_max_err = %f", iterations, G_max_err);
printf("\nThe total SET-UP time for MPI and boundary conditions is %lf", (t_setup_end-t_setup_strt));
printf("\nThe total time for SOLVING is %lf", (end-strt));
printf("\nThe total time for INDEPENDENT COMPUTE %lf", t_i_sum);
printf("\nThe total time for COMMUNICATION OVERHEAD is %lf", t_comm_sum);
printf("\nThe total time for MPI_ALLREDUCE() is %lf", t_allred_total);
}
MPI_Type_free(&x_subarray);
MPI_Type_free(&y_subarray);
MPI_Type_free(&z_subarray);
free(&old[0][0][0]);
free(&new[0][0][0]);
MPI_Finalize();
return 0;
}
P.S. : I am almost sure that the cost of spawning/waking the threads is not the reason for such a huge difference in the timing.
Please find attached Scalasca snapshot for INDEPENDENT COMPUTE of the Hybrid Program.
Using loop simd construct
#pragma omp parallel default(none) shared(old,new,PX,PY,PZ,l_max_err) private(i,j,k,diff)
{
#pragma omp for simd schedule(static) reduction(max:l_max_err)
for(i = 1; i <= PX ; i++)
{
for(j = 1; j<= PY; j++)
{
for(k = 1; k<= PZ; k++)
{
new[i][j][k] = (1/6.0) *(old[i-1][j][k] + old[i+1][j][k] + old[i][j-1][k] + old[i][j+1][k] + old[i][j][k-1] + old[i][j][k+1] );
diff = 1.0 - new[i][j][k];
diff = (diff > 0 ? diff : -1.0 * diff );
if(diff > l_max_err)
l_max_err = diff;
}
}
}
}
Solution 1:[1]
You frequently get memory access and cache issues when you just do one MPI process per socket on a CPU with multiple memory controllers. It can be on either the read or the write side, so you can't really say which. This is especially an issue when doing thread-parallel execution with lightweight compute tasks (e.g. math on arrays). One MPI process per socket in this case tends to fare significantly worse than pure MPI.
- In your BIOS, set up whatever the maximal NUMA per socket option is
- Use one MPI process per NUMA node.
- Try some different parameter values in schedule(static). I've rarely found the default to be best.
Essentially what this will do is ensure each bundle of threads only works on a single pool of memory.
Sources
This article follows the attribution requirements of Stack Overflow and is licensed under CC BY-SA 3.0.
Source: Stack Overflow
| Solution | Source |
|---|---|
| Solution 1 | user2560966 |