Commit aecabbb9 authored by lykos98's avatar lykos98
Browse files

added some work on lock free impl of H1, still not working as expected, poor performance

parent 5e0399c9

src/adp/adp.c

+40 −32
Original line number Diff line number Diff line 
#include "adp.h"

#include <unistd.h>



const border_t border_null = {.density = -1.0, .error = 0, .idx = NOBORDER};

const sparse_border_t sparse_border_null = {.density = -1.0, .error = 0, .idx = NOBORDER, .i = NOBORDER, .j = NOBORDER};
@@ -617,6 +618,8 @@ lock_t h1_lock_acquire(global_context_t* ctx, MPI_Win lock_window, int owner, id 
        while(result == LOCK_ACQUIRED && err == MPI_SUCCESS)

        {

            err = MPI_Compare_and_swap(&state, &compare, &result, MPI_LOCK_T, owner, pos, lock_window);

            MPI_Win_flush(owner, lock_window);

            usleep(100);

        }



        if(err != MPI_SUCCESS)
@@ -677,8 +680,7 @@ clusters_t Heuristic1(global_context_t *ctx) 
    MPI_Win_fence(0, win_datapoints);

    MPI_Win_lock_all(0,  win_datapoints);



#if defined(THREAD_FUNNELED)

#else

#if !defined(THREAD_FUNNELED)

    #pragma omp parallel for

#endif

    for(idx_t i = 0; i < n; ++i)
@@ -741,14 +743,17 @@ clusters_t Heuristic1(global_context_t *ctx) 
    MPI_Win_create(lock_array, n * sizeof(lock_t), sizeof(lock_t), MPI_INFO_NULL, ctx -> mpi_communicator, &win_locks);

    MPI_Win_fence(0, win_locks);



#ifdef EXP_H1

    MPI_Win_lock_all(0, win_to_remove_mask);

    MPI_Win_lock_all(0, win_locks);

#endif



#ifdef EXP_H1

    printf("Using experimental h1\n");

#endif





#if defined(THREAD_FUNNELED)

#else

    #pragma omp parallel for

#if !defined(THREAD_FUNNELED)

    #pragma omp parallel for schedule(dynamic)

#endif

    for(idx_t p = 0; p < n; ++p)

    {
@@ -769,7 +774,9 @@ clusters_t Heuristic1(global_context_t *ctx) 
                 *

                 * */



#ifdef EXPERIMENTAL_H1

#ifdef EXP_H1

                #pragma omp critical (h1_exp)

                {

                    int owner = foreign_owner(ctx, jidx);

                    idx_t jpos = jidx - ctx -> rank_idx_start[owner];
@@ -796,9 +803,9 @@ clusters_t Heuristic1(global_context_t *ctx) 
                    }



                    state = h1_lock_free(ctx, win_locks, owner, jpos, state);

                }

#else



                #pragma omp critical (h1_centers_elimination)

                #pragma omp critical (centers_elimination)                 

                {

                    int owner = foreign_owner(ctx, jidx);

                    idx_t jpos = jidx - ctx -> rank_idx_start[owner];
@@ -822,13 +829,16 @@ clusters_t Heuristic1(global_context_t *ctx) 
                    }



                    MPI_Win_unlock(owner, win_to_remove_mask);

                }

#endif

            }

        }

    }



#ifdef EXP_H1

    MPI_Win_unlock_all(win_to_remove_mask);

    MPI_Win_unlock_all(win_locks);

#endif

    

    MPI_Win_fence(0, win_to_remove_mask);

    MPI_Win_fence(0, win_locks);
@@ -1815,8 +1825,6 @@ void Heuristic3(global_context_t* ctx, clusters_t* cluster, float_t Z, int halo) 


        qsort(centers_dp, cluster -> centers.count, sizeof(datapoint_info_t), compare_dp_by_cidx);



        printf("Centers\n");



        master_finds_borders(ctx, cluster, Z, surviving_clusters, centers_dp);

        master_fixes_border_matrix_and_centers(ctx, cluster, Z, old_to_new, surviving_clusters, nclus);

        free(centers_dp);

src/main/main.c

+0 −47
Original line number Diff line number Diff line
@@ -232,52 +232,7 @@ void simulate_master_read_and_scatter(int dims, size_t n, global_context_t *ctx) 
    if (ctx->mpi_rank == 0) 

    {

        data = read_data_file(ctx,ctx -> input_data_file, ctx -> dims, ctx -> input_data_in_float32);

        

        //data = read_data_file(ctx, "../norm_data/50_blobs_more_var.npy", MY_TRUE);

        //ctx->dims = 2;

        //data = read_data_file(ctx, "../norm_data/blobs_small.npy", MY_FALSE);

        //data = read_data_file(ctx, "../norm_data/blobs_small.npy", MY_FALSE);

        // std_g0163178_Me14_091_0000

    

        // 100M points

        // 2D

        // std_g2980844_091_0000

        //data = read_data_file(ctx,"../norm_data/huge_blobs.npy",MY_FALSE);

        // 2B points

        // data = read_data_file(ctx,"../norm_data/very_huge_blobs.npy",MY_FALSE);

        // data = read_data_file(ctx,"../norm_data/hd_blobs.npy",5,MY_FALSE);

        

        //1B points

        // data = read_data_file(ctx,"../norm_data/eds_box_acc_normalized",5,MY_FALSE);

        // data = read_data_file(ctx,"../norm_data/eds_box_6d",6,MY_FALSE);



        // 190M points

        // std_g2980844_091_0000

        // data = read_data_file(ctx,"../norm_data/std_g2980844_091_0000",5,MY_TRUE);

        

        /* 1M points ca.*/

        //data = read_data_file(ctx,"../norm_data/std_LR_091_0001",5,MY_TRUE);



        /* BOX */

        // data = read_data_file(ctx,"../norm_data/std_Box_256_30_092_0000",MY_TRUE);



        /* 8M points */

        

        // data = read_data_file(ctx,"../norm_data/std_g0144846_Me14_091_0001",5,MY_TRUE);



        //88M 

        // data = read_data_file(ctx,"../norm_data/std_g5503149_091_0000",MY_TRUE);



        //

        //34 M

        //data = read_data_file(ctx,"../norm_data/std_g1212639_091_0001",MY_TRUE);

        

        //for weak scalability 

        //ctx->n_points = ctx->n_points / 4;

        //ctx->n_points = (ctx->n_points / 32) * ctx -> world_size;



        get_dataset_diagnostics(ctx, data);



    }

    

    /* communicate the total number of points*/
@@ -320,9 +275,7 @@ void simulate_master_read_and_scatter(int dims, size_t n, global_context_t *ctx) 
                int count_send = MIN(default_msg_len, send_counts[i] - already_sent_points); 

                MPI_Send(data + displacements[i] + already_sent_points, count_send, MPI_MY_FLOAT, i, ctx -> mpi_rank, ctx -> mpi_communicator);

                already_sent_points += count_send;

                //DB_PRINT("[RANK 0] has sent to rank %d %d elements out of %lu\n",i, already_sent_points, send_counts[i]);

            }

            //DB_PRINT("------------------------------------------------\n");

        }

    }

    else

src/tree/kdtreeV2.c

+0 −3
Original line number Diff line number Diff line
@@ -228,9 +228,6 @@ kdnode_v2* make_tree_kdnode_v2(kdnode_v2* t, int start, int end, kdnode_v2* pare 
        

    }

    

    





    int median_idx = -1;

    

    //if ((end - start) < 0) return 0;

src/tree/tree.c

+4 −3
Original line number Diff line number Diff line
@@ -1478,7 +1478,7 @@ void mpi_ngbh_search(global_context_t* ctx, datapoint_info_t* dp_info, top_kdtre 
    //ctx -> __local_heap_buffers = (heap_node*)MY_MALLOC(ctx -> local_n_points * k * sizeof(heap_node));

    MPI_Alloc_mem(ctx -> local_n_points * k * sizeof(heap_node), MPI_INFO_NULL, &(ctx -> __local_heap_buffers));



    #pragma omp parallel for

    #pragma omp parallel for schedule(dynamic)

    for(int p = 0; p < ctx -> local_n_points; ++p)

    {

        idx_t idx = local_tree -> _nodes[p].array_idx;
@@ -1663,7 +1663,7 @@ void mpi_ngbh_search(global_context_t* ctx, datapoint_info_t* dp_info, top_kdtre 
        //if(count_rcv_work_batches[p] > 0)

        {

            //heap_batches_per_node[p] = (heap_node*)MY_MALLOC(k * point_to_rcv_count[p] * sizeof(heap_node));

            #pragma omp parallel for

            #pragma omp parallel for schedule(dynamic)

            for(int batch = 0; batch < point_to_rcv_count[p]; ++batch)

            {

                heap H;
@@ -1822,7 +1822,7 @@ void mpi_ngbh_search(global_context_t* ctx, datapoint_info_t* dp_info, top_kdtre 
        }



        /* merge lists */

        #pragma omp paralell for

        #pragma omp parallel for

        for(int b = 0; b < ngbh_to_recv[rank_to_recv]; ++b)

        {

            int idx = local_idx_of_the_point[rank_to_recv][b];
@@ -1843,6 +1843,7 @@ void mpi_ngbh_search(global_context_t* ctx, datapoint_info_t* dp_info, top_kdtre 
        }







        MPI_Barrier(ctx -> mpi_communicator);

    }