model-net-mpi-replay.c

/*
 * Copyright (C) 2014 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */
#include <ross.h>
#include <inttypes.h>
#include <sys/stat.h>

#include "codes/codes-workload.h"
#include "codes/codes.h"
#include "codes/configuration.h"
#include "codes/codes_mapping.h"
#include "codes/model-net.h"
#include "codes/rc-stack.h"
#include "codes/quicklist.h"
#include "codes/quickhash.h"
#include "codes/codes-jobmap.h"

/* turning on track lp will generate a lot of output messages */
#define MN_LP_NM "modelnet_dragonfly_custom"
#define CONTROL_MSG_SZ 64
#define TRACE -1
#define MAX_WAIT_REQS 512
#define CS_LP_DBG 1
#define RANK_HASH_TABLE_SZ 2000
#define NW_LP_NM "nw-lp"
#define lprintf(_fmt, ...) \
        do {if (CS_LP_DBG) printf(_fmt, __VA_ARGS__);} while (0)
#define MAX_STATS 65536
#define PAYLOAD_SZ 1024

static int msg_size_hash_compare(
            void *key, struct qhash_head *link);

/* NOTE: Message tracking works in sequential mode only! */
static int debug_cols = 0;
/* Turning on this option slows down optimistic mode substantially. Only turn
 * on if you get issues with wait-all completion with traces. */
static int preserve_wait_ordering = 0;
static int enable_msg_tracking = 0;
static int is_synthetic = 0;
tw_lpid TRACK_LP = -1;

static double total_syn_data = 0;
static int unmatched = 0;
char workload_type[128];
char workload_file[8192];
char offset_file[8192];
static int wrkld_id;
static int num_net_traces = 0;
static int num_dumpi_traces = 0;
static int64_t EAGER_THRESHOLD = 8192;

static int alloc_spec = 0;
static tw_stime self_overhead = 10.0;
static tw_stime mean_interval = 100000;

/* Doing LP IO*/
static char lp_io_dir[256] = {'\0'};
static char sampling_dir[32] = {'\0'};
static lp_io_handle io_handle;
static unsigned int lp_io_use_suffix = 0;
static int do_lp_io = 0;

/* variables for loading multiple applications */
char workloads_conf_file[8192];
char alloc_file[8192];
int num_traces_of_job[5];
tw_stime soft_delay_mpi = 2500;
tw_stime nic_delay = 1000;
tw_stime copy_per_byte_eager = 0.55;
char file_name_of_job[5][8192];

struct codes_jobmap_ctx *jobmap_ctx;
struct codes_jobmap_params_list jobmap_p;

/* Variables for Cortex Support */
/* Matthieu's additions start */
#ifdef ENABLE_CORTEX_PYTHON
static char cortex_file[512] = "\0";
static char cortex_class[512] = "\0";
static char cortex_gen[512] = "\0";
#endif
/* Matthieu's additions end */

typedef struct nw_state nw_state;
typedef struct nw_message nw_message;
typedef int dumpi_req_id;

static int net_id = 0;
static float noise = 2.0;
static int num_nw_lps = 0, num_mpi_lps = 0;

static int num_syn_clients;

FILE * workload_log = NULL;
FILE * msg_size_log = NULL;
FILE * workload_agg_log = NULL;
FILE * workload_meta_log = NULL;

static uint64_t sample_bytes_written = 0;

long long num_bytes_sent=0;
long long num_bytes_recvd=0;

long long num_syn_bytes_sent = 0;
long long num_syn_bytes_recvd = 0;

double max_time = 0,  max_comm_time = 0, max_wait_time = 0, max_send_time = 0, max_recv_time = 0;
double avg_time = 0, avg_comm_time = 0, avg_wait_time = 0, avg_send_time = 0, avg_recv_time = 0;


/* runtime option for disabling computation time simulation */
static int disable_delay = 0;
static int enable_sampling = 0;
static double sampling_interval = 5000000;
static double sampling_end_time = 3000000000;
static int enable_debug = 0;

/* set group context */
struct codes_mctx group_ratio;

/* MPI_OP_GET_NEXT is for getting next MPI operation when the previous operation completes.
* MPI_SEND_ARRIVED is issued when a MPI message arrives at its destination (the message is transported by model-net and an event is invoked when it arrives.
* MPI_SEND_POSTED is issued when a MPI message has left the source LP (message is transported via model-net). */
enum MPI_NW_EVENTS
{
	MPI_OP_GET_NEXT=1,
	MPI_SEND_ARRIVED,
    MPI_SEND_ARRIVED_CB, // for tracking message times on sender
	MPI_SEND_POSTED,
    MPI_REND_ARRIVED,
    MPI_REND_ACK_ARRIVED,
    CLI_BCKGND_FIN,
    CLI_BCKGND_ARRIVE,
    CLI_BCKGND_GEN,
    CLI_NBR_FINISH,
};

struct mpi_workload_sample
{
    /* Sampling data */
    int nw_id;
    int app_id;
    unsigned long num_sends_sample;
    unsigned long num_bytes_sample;
    unsigned long num_waits_sample;
    double sample_end_time;
};
/* stores pointers of pending MPI operations to be matched with their respective sends/receives. */
struct mpi_msgs_queue
{
    int op_type;
    int tag;
    int source_rank;
    int dest_rank;
    int64_t num_bytes;
    tw_stime req_init_time;
	dumpi_req_id req_id;
    struct qlist_head ql;
};

/* stores request IDs of completed MPI operations (Isends or Irecvs) */
struct completed_requests
{
	int req_id;
    struct qlist_head ql;
    int index;
};

/* for wait operations, store the pending operation and number of completed waits so far. */
struct pending_waits
{
    int op_type;
    int req_ids[MAX_WAIT_REQS];
	int num_completed;
	int count;
    tw_stime start_time;
    struct qlist_head ql;
};

struct msg_size_info
{
    int64_t msg_size;
    int num_msgs;
    tw_stime agg_latency;
    tw_stime avg_latency;
    struct qhash_head * hash_link;
    struct qlist_head ql; 
};
typedef struct mpi_msgs_queue mpi_msgs_queue;
typedef struct completed_requests completed_requests;
typedef struct pending_waits pending_waits;

/* state of the network LP. It contains the pointers to send/receive lists */
struct nw_state
{
	long num_events_per_lp;
	tw_lpid nw_id;
	short wrkld_end;
    int app_id;
    int local_rank;

    int is_finished;
    int neighbor_completed;

    struct rc_stack * processed_ops;
    struct rc_stack * processed_wait_op;
    struct rc_stack * matched_reqs;
//    struct rc_stack * indices;

    /* count of sends, receives, collectives and delays */
	unsigned long num_sends;
	unsigned long num_recvs;
	unsigned long num_cols;
	unsigned long num_delays;
	unsigned long num_wait;
	unsigned long num_waitall;
	unsigned long num_waitsome;


	/* time spent by the LP in executing the app trace*/
	double start_time;

    double col_time;

    double reduce_time;
    int num_reduce;

    double all_reduce_time;
    int num_all_reduce;

	double elapsed_time;
	/* time spent in compute operations */
	double compute_time;
	/* time spent in message send/isend */
	double send_time;
	/* time spent in message receive */
	double recv_time;
	/* time spent in wait operation */
	double wait_time;
	/* FIFO for isend messages arrived on destination */
	struct qlist_head arrival_queue;
	/* FIFO for irecv messages posted but not yet matched with send operations */
	struct qlist_head pending_recvs_queue;
	/* List of completed send/receive requests */
	struct qlist_head completed_reqs;

    tw_stime cur_interval_end;
    
    /* Pending wait operation */
    struct pending_waits * wait_op;

    /* Message size latency information */
    struct qhash_table * msg_sz_table;
    struct qlist_head msg_sz_list;

    /* quick hash for maintaining message latencies */

    unsigned long num_bytes_sent;
    unsigned long num_bytes_recvd;

    unsigned long syn_data;
    unsigned long gen_data;
    
    /* For sampling data */
    int sampling_indx;
    int max_arr_size;
    struct mpi_workload_sample * mpi_wkld_samples;
    char output_buf[512];
    char col_stats[64];
};

/* data for handling reverse computation.
* saved_matched_req holds the request ID of matched receives/sends for wait operations.
* ptr_match_op holds the matched MPI operation which are removed from the queues when a send is matched with the receive in forward event handler.
* network event being sent. op is the MPI operation issued by the network workloads API. rv_data holds the data for reverse computation (TODO: Fill this data structure only when the simulation runs in optimistic mode). */
struct nw_message
{
   // forward message handler
   int msg_type;
   int op_type;
   model_net_event_return event_rc;

   struct
   {
       tw_lpid src_rank;
       int dest_rank;
       int64_t num_bytes;
       int num_matched;
       int data_type;
       double sim_start_time;
       // for callbacks - time message was received
       double msg_send_time;
       int req_id;
       int matched_req;
       int tag;
       int app_id;
       int found_match;
       short wait_completed;
       short rend_send;
   } fwd;
   struct
   {
       double saved_send_time;
       double saved_recv_time;
       double saved_wait_time;
       double saved_delay;
       int64_t saved_num_bytes;
   } rc;
};

static void send_ack_back(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, mpi_msgs_queue * mpi_op, int matched_req);

static void send_ack_back_rc(nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* executes MPI isend and send operations */
static void codes_exec_mpi_send(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op, int is_rend);
/* execute MPI irecv operation */
static void codes_exec_mpi_recv(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp, struct codes_workload_op * mpi_op);
/* reverse of mpi recv function. */
static void codes_exec_mpi_recv_rc(
        nw_state* s, tw_bf * bf, nw_message* m, tw_lp* lp);
/* execute the computational delay */
static void codes_exec_comp_delay(
        nw_state* s, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op);
/* gets the next MPI operation from the network-workloads API. */
static void get_next_mpi_operation(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* reverse handler of get next mpi operation. */
static void get_next_mpi_operation_rc(
        nw_state* s, tw_bf * bf, nw_message * m, tw_lp * lp);
/* Makes a call to get_next_mpi_operation. */
static void codes_issue_next_event(tw_lp* lp);
/* reverse handler of next operation */
static void codes_issue_next_event_rc(tw_lp* lp);


///////////////////// HELPER FUNCTIONS FOR MPI MESSAGE QUEUE HANDLING ///////////////
/* upon arrival of local completion message, inserts operation in completed send queue */
/* upon arrival of an isend operation, updates the arrival queue of the network */
static void update_completed_queue(
        nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp, dumpi_req_id req_id);
/* reverse of the above function */
static void update_completed_queue_rc(
        nw_state*s,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp);
static void update_arrival_queue(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse of the above function */
static void update_arrival_queue_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* callback to a message sender for computing message time */
static void update_message_time(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);
/* reverse for computing message time */
static void update_message_time_rc(
        nw_state*s, tw_bf* bf, nw_message* m, tw_lp * lp);

/* conversion from seconds to eanaoseconds */
static tw_stime s_to_ns(tw_stime ns);

/*static void update_message_size_rc(
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{*/
/*TODO: Complete reverse handler */
/*    (void)ns;
    (void)lp;
    (void)bf;
    (void)m;
}*/
/* update the message size */
static void update_message_size(
        struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m,
        mpi_msgs_queue * qitem,
        int is_eager,
        int is_send)
{
            (void)bf;
            (void)is_eager;

            struct qhash_head * hash_link = NULL;
            tw_stime msg_init_time = qitem->req_init_time;
        
            if(!ns->msg_sz_table)
                ns->msg_sz_table = qhash_init(msg_size_hash_compare, quickhash_64bit_hash, RANK_HASH_TABLE_SZ); 
            
            hash_link = qhash_search(ns->msg_sz_table, &(qitem->num_bytes));

            if(is_send)
                msg_init_time = m->fwd.sim_start_time;
            
            /* update hash table */
            if(!hash_link)
            {
                struct msg_size_info * msg_info = (struct msg_size_info*)malloc(sizeof(struct msg_size_info));
                msg_info->msg_size = qitem->num_bytes;
                msg_info->num_msgs = 1;
                msg_info->agg_latency = tw_now(lp) - msg_init_time;
                msg_info->avg_latency = msg_info->agg_latency;
                qhash_add(ns->msg_sz_table, &(msg_info->msg_size), msg_info->hash_link);
                qlist_add(&msg_info->ql, &ns->msg_sz_list);
                //printf("\n Msg size %d aggregate latency %f num messages %d ", m->fwd.num_bytes, msg_info->agg_latency, msg_info->num_msgs);
            }
            else
            {
                struct msg_size_info * tmp = qhash_entry(hash_link, struct msg_size_info, hash_link);
                tmp->num_msgs++;
                tmp->agg_latency += tw_now(lp) - msg_init_time;  
                tmp->avg_latency = (tmp->agg_latency / tmp->num_msgs);
//                printf("\n Msg size %d aggregate latency %f num messages %d ", qitem->num_bytes, tmp->agg_latency, tmp->num_msgs);
            }
}
static void notify_background_traffic_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
    (void)ns;
    (void)bf;
    (void)m;
        
    int num_jobs = codes_jobmap_get_num_jobs(jobmap_ctx); 
    
    for(int i = 0; i < num_jobs - 1; i++)
        tw_rand_reverse_unif(lp->rng); 
}

static void notify_background_traffic(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
        (void)bf;
        (void)m;

        struct codes_jobmap_id jid; 
        jid = codes_jobmap_to_local_id(ns->nw_id, jobmap_ctx);
        
        int num_jobs = codes_jobmap_get_num_jobs(jobmap_ctx); 
        
        for(int other_id = 0; other_id < num_jobs; other_id++)
        {
            if(other_id == jid.job)
                continue;

            struct codes_jobmap_id other_jid;
            other_jid.job = other_id;

            int num_other_ranks = codes_jobmap_get_num_ranks(other_id, jobmap_ctx);

            lprintf("\n Other ranks %d ", num_other_ranks);
            tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, mean_interval/10000);
            tw_lpid global_dest_id;
     
            for(int k = 0; k < num_other_ranks; k++)    
            {
                other_jid.rank = k;
                int intm_dest_id = codes_jobmap_to_global_id(other_jid, jobmap_ctx); 
                global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);

                tw_event * e;
                struct nw_message * m_new;  
                e = tw_event_new(global_dest_id, ts, lp);
                m_new = (struct nw_message*)tw_event_data(e);
                m_new->msg_type = CLI_BCKGND_FIN;
                tw_event_send(e);   
            }
        }
        return;
}
static void notify_neighbor_rc(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
       if(bf->c0)
       {
            notify_background_traffic_rc(ns, lp, bf, m);
            return;
       }
   
       if(bf->c1)
       {
          tw_rand_reverse_unif(lp->rng); 
       }
} 
static void notify_neighbor(
	    struct nw_state * ns,
        tw_lp * lp,
        tw_bf * bf,
        struct nw_message * m)
{
    if(ns->local_rank == num_dumpi_traces - 1 
            && ns->is_finished == 1
            && ns->neighbor_completed == 1)
    {
//        printf("\n All workloads completed, notifying background traffic ");
        bf->c0 = 1;
        notify_background_traffic(ns, lp, bf, m);
        return;
    }
    
    struct codes_jobmap_id nbr_jid;
    nbr_jid.job = ns->app_id;
    tw_lpid global_dest_id;

    if(ns->is_finished == 1 && (ns->neighbor_completed == 1 || ns->local_rank == 0))
    {
        bf->c1 = 1;

//        printf("\n Local rank %d notifying neighbor %d ", ns->local_rank, ns->local_rank+1);
        tw_stime ts = (1.1 * g_tw_lookahead) + tw_rand_exponential(lp->rng, mean_interval/10000);
        nbr_jid.rank = ns->local_rank + 1;
        
        /* Send a notification to the neighbor about completion */
        int intm_dest_id = codes_jobmap_to_global_id(nbr_jid, jobmap_ctx); 
        global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);
       
        tw_event * e;
        struct nw_message * m_new;  
        e = tw_event_new(global_dest_id, ts, lp);
        m_new = (struct nw_message*)tw_event_data(e); 
        m_new->msg_type = CLI_NBR_FINISH;
        tw_event_send(e);   
    }
}
void finish_bckgnd_traffic_rc(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
        (void)b;
        (void)msg;
        (void)lp;

        ns->is_finished = 0;
        return;
}
void finish_bckgnd_traffic(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
        (void)b;
        (void)msg;
        ns->is_finished = 1;
        lprintf("\n LP %llu completed sending data %lu completed at time %lf ", LLU(lp->gid), ns->gen_data, tw_now(lp));
        return;
}

void finish_nbr_wkld_rc(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
    ns->neighbor_completed = 0;
    
    notify_neighbor_rc(ns, lp, b, msg);
}

void finish_nbr_wkld(
    struct nw_state * ns,
    tw_bf * b,
    struct nw_message * msg,
    tw_lp * lp)
{
    ns->neighbor_completed = 1;

    notify_neighbor(ns, lp, b, msg);
}
static void gen_synthetic_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    if(bf->c0)
        return;

    model_net_event_rc2(lp, &m->event_rc);
    s->gen_data -= PAYLOAD_SZ;

    num_syn_bytes_sent -= PAYLOAD_SZ;
    tw_rand_reverse_unif(lp->rng);
    tw_rand_reverse_unif(lp->rng);

}

/* generate synthetic traffic */
static void gen_synthetic_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    if(s->is_finished == 1)
    {
        bf->c0 = 1;
        return;
    }

    /* Get job information */
    tw_lpid global_dest_id;

    struct codes_jobmap_id jid;
    jid = codes_jobmap_to_local_id(s->nw_id, jobmap_ctx); 

    int num_clients = codes_jobmap_get_num_ranks(jid.job, jobmap_ctx);
    int dest_svr = tw_rand_integer(lp->rng, 0, num_clients - 1);

    if(dest_svr == s->local_rank)
    {
       dest_svr = (s->local_rank + 1) % num_clients;
    }
   
    jid.rank = dest_svr;

    int intm_dest_id = codes_jobmap_to_global_id(jid, jobmap_ctx); 
    global_dest_id = codes_mapping_get_lpid_from_relative(intm_dest_id, NULL, NW_LP_NM, NULL, 0);

    nw_message remote_m;
    remote_m.fwd.sim_start_time = tw_now(lp);
    remote_m.fwd.dest_rank = dest_svr;
    remote_m.msg_type = CLI_BCKGND_ARRIVE;
    remote_m.fwd.num_bytes = PAYLOAD_SZ;
    remote_m.fwd.app_id = s->app_id;
    remote_m.fwd.src_rank = s->local_rank;

    m->event_rc = model_net_event(net_id, "synthetic-tr", global_dest_id, PAYLOAD_SZ, 0.0, 
            sizeof(nw_message), (const void*)&remote_m, 
            0, NULL, lp);
    
    s->gen_data += PAYLOAD_SZ;
    num_syn_bytes_sent += PAYLOAD_SZ; 

    /* New event after MEAN_INTERVAL */  
    tw_stime ts = mean_interval  + tw_rand_exponential(lp->rng, noise); 
    tw_event * e;
    nw_message * m_new;
    e = tw_event_new(lp->gid, ts, lp);
    m_new = (struct nw_message*)tw_event_data(e);
    m_new->msg_type = CLI_BCKGND_GEN;
    tw_event_send(e);
}

void arrive_syn_tr_rc(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    (void)bf;
    (void)m;
    (void)lp;
//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    int data = m->fwd.num_bytes;
    s->syn_data -= data;
    num_syn_bytes_recvd -= data;
}
void arrive_syn_tr(nw_state * s, tw_bf * bf, nw_message * m, tw_lp * lp)
{
    (void)bf;
    (void)lp;

//    printf("\n Data arrived %d total data %ld ", m->fwd.num_bytes, s->syn_data);
    int data = m->fwd.num_bytes;
    s->syn_data += data;
    num_syn_bytes_recvd += data;
}
/* Debugging functions, may generate unused function warning */
/*static void print_waiting_reqs(uint32_t * reqs, int count)
{
    lprintf("\n Waiting reqs: %d count", count);
    int i;
    for(i = 0; i < count; i++ )
        lprintf(" %d ", reqs[i]);
}*/
static void print_msgs_queue(struct qlist_head * head, int is_send)
{
    if(is_send)
        printf("\n Send msgs queue: ");
    else
        printf("\n Recv msgs queue: ");

    struct qlist_head * ent = NULL;
    mpi_msgs_queue * current = NULL;
    qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, mpi_msgs_queue, ql);
            printf(" \n Source %d Dest %d bytes %"PRId64" tag %d ", current->source_rank, current->dest_rank, current->num_bytes, current->tag);
       }
}
/*static void print_completed_queue(tw_lp * lp, struct qlist_head * head)
{
//    printf("\n Completed queue: ");
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      tw_output(lp, "\n");
      qlist_for_each(ent, head)
       {
            current = qlist_entry(ent, completed_requests, ql);
            tw_output(lp, " %llu ", current->req_id);
       }
}*/
static int clear_completed_reqs(nw_state * s,
        tw_lp * lp,
        int * reqs, int count)
{
    (void)s;
    (void)lp;

    int i, matched = 0;

    for( i = 0; i < count; i++)
    {
      struct qlist_head * ent = NULL;
      struct completed_requests * current = NULL;
      struct completed_requests * prev = NULL;

      int index = 0;
      qlist_for_each(ent, &s->completed_reqs)
       {
           if(prev)
           {
              rc_stack_push(lp, prev, free, s->matched_reqs);
              prev = NULL;
           }
            
           current = qlist_entry(ent, completed_requests, ql);
           current->index = index; 
            if(current->req_id == reqs[i])
            {
                ++matched;
                qlist_del(&current->ql);
                prev = current;
            }
            ++index;
       }

      if(prev)
      {
         rc_stack_push(lp, prev, free, s->matched_reqs);
         prev = NULL;
      }
    }
    return matched;
}
static void add_completed_reqs(nw_state * s,
        tw_lp * lp,
        int count)
{
    (void)lp;
    for(int i = 0; i < count; i++)
    {
       struct completed_requests * req = (struct completed_requests*)rc_stack_pop(s->matched_reqs);

       // turn on only if wait-all unmatched error arises in optimistic mode.
       if(preserve_wait_ordering)
       {
            if(req->index == 0)
            {
                qlist_add(&req->ql, &s->completed_reqs);
            }
            else
            {
                int index = 1;
                struct qlist_head * ent = NULL;
                qlist_for_each(ent, &s->completed_reqs) 
                {
                    if(index == req->index)
                    {
                        qlist_add(&req->ql, ent);
                    }
                }//end qlist
            }// end else*/
       }
       else
       {
               qlist_add(&req->ql, &s->completed_reqs);
       }
    }//end for
}

/* helper function - maps an MPI rank to an LP id */
static tw_lpid rank_to_lpid(int rank)
{
    return codes_mapping_get_lpid_from_relative(rank, NULL, "nw-lp", NULL, 0);
}

static int notify_posted_wait(nw_state* s,
        tw_bf * bf, nw_message * m, tw_lp * lp,
        int completed_req)
{
    (void)bf;

    struct pending_waits* wait_elem = s->wait_op;
    int wait_completed = 0;

    m->fwd.wait_completed = 0;

    if(!wait_elem)
        return 0;

    int op_type = wait_elem->op_type;

    if(op_type == CODES_WK_WAIT &&
            (wait_elem->req_ids[0] == completed_req))
    {
            m->fwd.wait_completed = 1;
            wait_completed = 1;
    }
    else if(op_type == CODES_WK_WAITALL
            || op_type == CODES_WK_WAITANY
            || op_type == CODES_WK_WAITSOME)
    {
        int i;
        for(i = 0; i < wait_elem->count; i++)
        {
            if(wait_elem->req_ids[i] == completed_req)
            {
                wait_elem->num_completed++;
                if(wait_elem->num_completed > wait_elem->count)
                    printf("\n Num completed %d count %d LP %llu ",
                            wait_elem->num_completed,
                            wait_elem->count,
                            LLU(lp->gid));
//                if(wait_elem->num_completed > wait_elem->count)
//                    tw_lp_suspend(lp, 1, 0);

                if(wait_elem->num_completed >= wait_elem->count)
                {
                    if(enable_debug)
                        fprintf(workload_log, "\n(%lf) APP ID %d MPI WAITALL COMPLETED AT %llu ", tw_now(lp), s->app_id, LLU(s->nw_id));
                    wait_completed = 1;
                }

                m->fwd.wait_completed = 1;
            }
        }
    }
    return wait_completed;
}

/* reverse handler of MPI wait operation */
static void codes_exec_mpi_wait_rc(nw_state* s, tw_bf * bf, tw_lp* lp, nw_message * m)
{
   if(bf->c1)
    {
        completed_requests * qi = (completed_requests*)rc_stack_pop(s->processed_ops);
        if(m->fwd.found_match == 0)
        {
            qlist_add(&qi->ql, &s->completed_reqs);
        }
        else
        {
           int index = 1;
           struct qlist_head * ent = NULL;
           qlist_for_each(ent, &s->completed_reqs)
           {
                if(index == m->fwd.found_match)
                {
                    qlist_add(&qi->ql, ent);
                    break;
                }
                index++;
           }
        }
        codes_issue_next_event_rc(lp);
        return;
    }
         struct pending_waits * wait_op = s->wait_op;
         free(wait_op);
         s->wait_op = NULL;
}

/* execute MPI wait operation */
static void codes_exec_mpi_wait(nw_state* s, tw_bf * bf, nw_message * m, tw_lp* lp, struct codes_workload_op * mpi_op)
{
    /* check in the completed receives queue if the request ID has already been completed.*/
    assert(!s->wait_op);
    int req_id = mpi_op->u.wait.req_id;

    struct completed_requests* current = NULL;

    struct qlist_head * ent = NULL;
    int index = 0;
    qlist_for_each(ent, &s->completed_reqs)
    {
        current = qlist_entry(ent, completed_requests, ql);
        if(current->req_id == req_id)
        {
            bf->c1=1;
            qlist_del(&current->ql);
            rc_stack_push(lp, current, free, s->processed_ops);
            codes_issue_next_event(lp);
            m->fwd.found_match = index;
            /*if(s->nw_id == (tw_lpid)TRACK_LP)
            {
                tw_output(lp, "\n wait matched at post %d ", req_id);
                print_completed_queue(lp, &s->completed_reqs);
            }*/
            return;
        }
        ++index;
    }

    /*if(s->nw_id == (tw_lpid)TRACK_LP)
    {
        tw_output(lp, "\n wait posted %llu ", req_id);
        print_completed_queue(lp, &s->completed_reqs);
    }*/
    /* If not, add the wait operation in the pending 'waits' list. */
    struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
    wait_op->op_type = mpi_op->op_type;
    wait_op->req_ids[0] = req_id;
    wait_op->count = 1;
    wait_op->num_completed = 0;
    wait_op->start_time = tw_now(lp);
    s->wait_op = wait_op;

    return;
}

static void codes_exec_mpi_wait_all_rc(
        nw_state* s,
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp)
{
  if(bf->c1)
  {
    int sampling_indx = s->sampling_indx;
    s->mpi_wkld_samples[sampling_indx].num_waits_sample--;

    if(bf->c2)
    {
        s->cur_interval_end -= sampling_interval;
        s->sampling_indx--;
    }
  }
  if(s->wait_op)
  {
      struct pending_waits * wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
  }
  else
  {
      add_completed_reqs(s, lp, m->fwd.num_matched);
      codes_issue_next_event_rc(lp);
  }
  return;
}

static void codes_exec_mpi_wait_all(
        nw_state* s,
        tw_bf * bf,
        nw_message * m,
        tw_lp* lp,
        struct codes_workload_op * mpi_op)
{
  if(enable_debug)
    fprintf(workload_log, "\n MPI WAITALL POSTED AT %llu ", LLU(s->nw_id));

  if(enable_sampling)
  {
    bf->c1 = 1;
    if(tw_now(lp) >= s->cur_interval_end)
    {
        bf->c2 = 1;
        int indx = s->sampling_indx;
        s->mpi_wkld_samples[indx].nw_id = s->nw_id;
        s->mpi_wkld_samples[indx].app_id = s->app_id;
        s->mpi_wkld_samples[indx].sample_end_time = s->cur_interval_end;
        s->cur_interval_end += sampling_interval;
        s->sampling_indx++;
    }
    if(s->sampling_indx >= MAX_STATS)
    {
        struct mpi_workload_sample * tmp = (struct mpi_workload_sample*)calloc((MAX_STATS + s->max_arr_size), sizeof(struct mpi_workload_sample));
        memcpy(tmp, s->mpi_wkld_samples, s->sampling_indx);
        free(s->mpi_wkld_samples);
        s->mpi_wkld_samples = tmp;
        s->max_arr_size += MAX_STATS;
    }
    int indx = s->sampling_indx;
    s->mpi_wkld_samples[indx].num_waits_sample++;
  }
  int count = mpi_op->u.waits.count;
  /* If the count is not less than max wait reqs then stop */
  assert(count < MAX_WAIT_REQS);

  int i = 0, num_matched = 0;
  m->fwd.num_matched = 0;

  /*if(lp->gid == TRACK_LP)
  {
      printf("\n MPI Wait all posted ");
      print_waiting_reqs(mpi_op->u.waits.req_ids, count);
      print_completed_queue(lp, &s->completed_reqs);
  }*/
      /* check number of completed irecvs in the completion queue */
  for(i = 0; i < count; i++)
  {
      int req_id = mpi_op->u.waits.req_ids[i];
      struct qlist_head * ent = NULL;
      struct completed_requests* current = NULL;
      qlist_for_each(ent, &s->completed_reqs)
       {
            current = qlist_entry(ent, struct completed_requests, ql);
            if(current->req_id == req_id)
                num_matched++;
       }
  }

  m->fwd.found_match = num_matched;
  if(num_matched == count)
  {
    /* No need to post a MPI Wait all then, issue next event */
      /* Remove all completed requests from the list */
      m->fwd.num_matched = clear_completed_reqs(s, lp, mpi_op->u.waits.req_ids, count);
      struct pending_waits* wait_op = s->wait_op;
      free(wait_op);
      s->wait_op = NULL;
      codes_issue_next_event(lp);
  }
  else
  {
      /* If not, add the wait operation in the pending 'waits' list. */
	  struct pending_waits* wait_op = (struct pending_waits*)malloc(sizeof(struct pending_waits));
	  wait_op->count = count;
      wait_op->op_type = mpi_op->op_type;
      assert(count < MAX_WAIT_REQS);

      for(i = 0; i < count; i++)
          wait_op->req_ids[i] =  mpi_op->u.waits.req_ids[i];

	  wait_op->num_completed = num_matched;
	  wait_op->start_time = tw_now(lp);
      s->wait_op = wait_op;
  }
  return;
}

/* search for a matching mpi operation and remove it from the list.
 * Record the index in the list from where the element got deleted.
 * Index is used for inserting the element once again in the queue for reverse computation. */
static int rm_matching_rcv(nw_state * ns,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp,
        mpi_msgs_queue * qitem)
{
    int matched = 0;
    int index = 0;
    int is_rend = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

    qlist_for_each(ent, &ns->pending_recvs_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
        if(//(qi->num_bytes == qitem->num_bytes)
                //&& 
               ((qi->tag == qitem->tag) || qi->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qi->source_rank == -1))
        {
            matched = 1;
            qi->num_bytes = qitem->num_bytes;
            break;
        }
        ++index;
    }

    if(matched)
    {
        if(enable_msg_tracking && qitem->num_bytes < EAGER_THRESHOLD)
        {
            update_message_size(ns, lp, bf, m, qitem, 1, 1);
        }
        if(qitem->num_bytes >= EAGER_THRESHOLD)
        {
            /* Matching receive found, need to notify the sender to transmit
             * the data * (only works in sequential mode)*/
            bf->c10 = 1;
            is_rend = 1;
            send_ack_back(ns, bf, m, lp, qitem, qi->req_id);
        }
        else
        {
            m->rc.saved_recv_time = ns->recv_time;
            ns->recv_time += (tw_now(lp) - m->fwd.sim_start_time);
        }
        if(qi->op_type == CODES_WK_IRECV && !is_rend)
        {
            bf->c9 = 1;
            /*if(ns->nw_id == (tw_lpid)TRACK_LP)
            {
                printf("\n Completed irecv req id %d ", qi->req_id);
            }*/
            update_completed_queue(ns, bf, m, lp, qi->req_id);
        }
        else if(qi->op_type == CODES_WK_RECV && !is_rend)
        {
            bf->c8 = 1;
            codes_issue_next_event(lp);
        }

        qlist_del(&qi->ql);

        rc_stack_push(lp, qi, free, ns->processed_ops);
        return index;
    }
    return -1;
}

static int rm_matching_send(nw_state * ns,
        tw_bf * bf,
        nw_message * m,
        tw_lp * lp, mpi_msgs_queue * qitem)
{
    int matched = 0;
    struct qlist_head *ent = NULL;
    mpi_msgs_queue * qi = NULL;

    int index = 0;
    qlist_for_each(ent, &ns->arrival_queue){
        qi = qlist_entry(ent, mpi_msgs_queue, ql);
        if(//(qi->num_bytes == qitem->num_bytes) // it is not a requirement in MPI that the send and receive sizes match
                // && 
		(qi->tag == qitem->tag || qitem->tag == -1)
                && ((qi->source_rank == qitem->source_rank) || qitem->source_rank == -1))
        {
            qitem->num_bytes = qi->num_bytes;
            matched = 1;