コード例 #1
0
ファイル: mtl_portals_recv.c プロジェクト: aosm/openmpi
static void 
ompi_mtl_portals_wait_for_put_end(ptl_seq_t link)
{
    ptl_event_t ev;
    int ret;

    /* wait for a PUT_END event that matches the message we're looking for */
    while (true) {
        ret = PtlEQWait(ompi_mtl_portals.ptl_unexpected_recv_eq_h,&ev);
        if (PTL_OK == ret) {
            if (PTL_EVENT_PUT_START == ev.type) {
                ompi_free_list_item_t *item;
                ompi_mtl_portals_event_t *recv_event;

                OMPI_FREE_LIST_GET(&ompi_mtl_portals.event_fl, item, ret);
                recv_event = (ompi_mtl_portals_event_t*) item;
                recv_event->ev = ev;
                recv_event->is_complete = false;
                opal_list_append(&(ompi_mtl_portals.unexpected_messages),
                                 (opal_list_item_t*) recv_event);

               if (PTL_IS_SHORT_MSG(recv_event->ev.match_bits)) {
                    ompi_mtl_portals_recv_short_block_t *block =
                        recv_event->ev.md.user_ptr;
                    OPAL_THREAD_ADD32(&block->pending, 1);
               }

            } else if (PTL_EVENT_PUT_END == ev.type) {
                if (link == ev.link) {
                    /* the one we want */
                    return;
                }
                /* otherwise match it up */
                ompi_mtl_portals_match_up_put_end(ev.link);
            } else {
                opal_output(fileno(stderr)," Unrecognised event type - %d - ompi_mtl_portals_wait_for_put_end : %d \n",ev.type,ret);
                abort(); 
            }
        } else {
            opal_output(fileno(stderr)," Error returned in ompi_mtl_portals_wait_for_put_end from PtlEQWait : %d \n",ret);
            abort();
        }
    }
}
コード例 #2
0
ファイル: clone.c プロジェクト: arnolda/scafacos
int
server(void *arg)
{

       int ret;
       int num_interfaces;
       ptl_handle_ni_t nih;
       ptl_handle_eq_t eqh;
       ptl_ni_limits_t ptl_limits;
       ptl_event_t ev_t;
       ptl_event_t *ev=&ev_t;


       printf("IN SERVER\n");

       if ((ret=PtlInit(&num_interfaces)) != PTL_OK) {
               printf("%s: PtlInit failed: %d\n", __FUNCTION__, ret);
               exit(1);
       }
       printf("%s: PtlInit succeeds (%d:%d)\n", __FUNCTION__, ret, num_interfaces);

       if (((ret=PtlNIInit(
               IFACE_FROM_BRIDGE_AND_NALID(PTL_BRIDGE_UK, PTL_IFACE_SS),
               PTL_PID_ANY, NULL, &ptl_limits, &nih)) != PTL_OK) && (ret != PTL_IFACE_DUP)) {
               printf("%s: PtlNIInit failed: %d\n", __FUNCTION__, ret);
               exit(1);
       }
       printf("%s: PtlNIInit succeeds (%d)\n", __FUNCTION__, ret);

       if ((ret=PtlEQAlloc(nih, 4096, NULL, &eqh)) != PTL_OK) {
               printf("%s: PtlEQAlloc failed: %d\n",
                       __FUNCTION__, ret);
               exit(1);
       }
       iv=11;
       ret=PtlEQWait(nih,ev);
       printf("%s: PtlEQAlloc succeeds\n", __FUNCTION__);
       printf("%d\n",iv);
       iv=13;

       while(1);
}
コード例 #3
0
ファイル: events_hotpotato.c プロジェクト: foool/portals4
int main(int   argc,
         char *argv[])
{
    ptl_handle_ni_t ni_logical;
    ptl_pt_index_t  logical_pt_index;
    ptl_process_t   myself;
    struct timeval  start, stop;
    int             potato = 0;
    ENTRY_T         potato_catcher;
    HANDLE_T        potato_catcher_handle;
    ptl_md_t        potato_launcher;
    ptl_handle_md_t potato_launcher_handle;
    int             num_procs;
    ptl_handle_eq_t pt_eq_handle;

    CHECK_RETURNVAL(PtlInit());

    CHECK_RETURNVAL(libtest_init());

    num_procs = libtest_get_size();

    if (NULL != getenv("MAKELEVEL") && num_procs > 2) {
        return 77;
    }

    CHECK_RETURNVAL(PtlNIInit
                        (PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL, PTL_PID_ANY,
                        NULL, NULL, &ni_logical));

    CHECK_RETURNVAL(PtlSetMap(ni_logical, num_procs,
                              libtest_get_mapping(ni_logical)));

    CHECK_RETURNVAL(PtlGetId(ni_logical, &myself));
    CHECK_RETURNVAL(PtlEQAlloc(ni_logical, 100, &pt_eq_handle));
    CHECK_RETURNVAL(PtlPTAlloc
                        (ni_logical, 0, pt_eq_handle, PTL_PT_ANY,
                        &logical_pt_index));
    assert(logical_pt_index == 0);
    /* Now do the initial setup on ni_logical */
    potato_catcher.start   = &potato;
    potato_catcher.length  = sizeof(potato);
    potato_catcher.uid     = PTL_UID_ANY;
    potato_catcher.options = OPTIONS;
#if INTERFACE == 1
    potato_catcher.match_id.rank = PTL_RANK_ANY;
    potato_catcher.match_bits    = 1;
    potato_catcher.ignore_bits   = ~potato_catcher.match_bits;
#endif
    potato_catcher.ct_handle = PTL_CT_NONE;
    CHECK_RETURNVAL(APPEND
                        (ni_logical, logical_pt_index, &potato_catcher,
                        PTL_PRIORITY_LIST, NULL, &potato_catcher_handle));
    {
        ptl_event_t event;
        CHECK_RETURNVAL(PtlEQWait(pt_eq_handle, &event));       // wait for link event
        assert(event.type == PTL_EVENT_LINK);
    }
    /* Now do a barrier (on ni_physical) to make sure that everyone has their
     * logical interface set up */
    libtest_barrier();

    /* now I can communicate between ranks with ni_logical */

    /* set up the potato launcher */
    potato_launcher.start     = &potato;
    potato_launcher.length    = sizeof(potato);
    potato_launcher.options   = PTL_MD_EVENT_CT_ACK | PTL_MD_EVENT_CT_SEND;
    potato_launcher.eq_handle = PTL_EQ_NONE;    // i.e. don't queue send events
    CHECK_RETURNVAL(PtlCTAlloc(ni_logical, &potato_launcher.ct_handle));
    CHECK_RETURNVAL(PtlMDBind
                        (ni_logical, &potato_launcher, &potato_launcher_handle));

    /* rank 0 starts the potato going */
    if (myself.rank == 0) {
        ptl_process_t nextrank;
        nextrank.rank  = myself.rank + 1;
        nextrank.rank *= (nextrank.rank <= num_procs - 1);
        gettimeofday(&start, NULL);
        CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0, potato_launcher.length,
                               (LOOPS == 1) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
                               nextrank, logical_pt_index, 1, 0, NULL, 1));
    }

    {                                  /* the potato-passing loop */
        size_t         waitfor;
        ptl_ct_event_t ctc;
        ptl_process_t  nextrank;
        nextrank.rank  = myself.rank + 1;
        nextrank.rank *= (nextrank.rank <= num_procs - 1);
        for (waitfor = 1; waitfor <= LOOPS; ++waitfor) {
            do {
                ptl_event_t event;
                CHECK_RETURNVAL(PtlEQWait(pt_eq_handle, &event));       // wait for potato
                if (event.type != PTL_EVENT_PUT) {
                    printf("unexpected event: %i\n", (int)event.type);
                } else {
                    break;
                }
            } while (1);
            /* I have the potato! */
            ++potato;
            if (potato < LOOPS * (num_procs)) { // otherwise, the recipient may have exited
                /* Bomb's away! */
                if (myself.rank == 0) {
                    CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0,
                                           potato_launcher.length,
                                           (waitfor == (LOOPS - 1)) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
                                           nextrank, logical_pt_index, 3, 0, NULL, 2));
                } else {
                    CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0,
                                           potato_launcher.length,
                                           (waitfor == LOOPS) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
                                           nextrank, logical_pt_index, 3, 0, NULL, 2));
                }
            }
        }
        // make sure that last send completed before exiting
        CHECK_RETURNVAL(PtlCTWait(potato_launcher.ct_handle, LOOPS + 1, &ctc));
        assert(ctc.failure == 0);
        if (myself.rank == 0) {
            printf("Final value of potato = %i\n", potato);
        }
    }
    if (myself.rank == 0) {
        double accumulate = 0.0;
        gettimeofday(&stop, NULL);
        accumulate =
            (stop.tv_sec + stop.tv_usec * 1e-6) - (start.tv_sec +
                                                   start.tv_usec * 1e-6);
        /* calculate the average time waiting */
        printf("Total time: %g secs\n", accumulate);
        accumulate /= LOOPS;
        printf("Average time around the loop: %g microseconds\n",
               accumulate * 1e6);
        accumulate /= num_procs;
        printf("Average catch-to-toss latency: %g microseconds\n",
               accumulate * 1e6);
    }

    /* cleanup */
    CHECK_RETURNVAL(PtlMDRelease(potato_launcher_handle));
    CHECK_RETURNVAL(PtlCTFree(potato_launcher.ct_handle));
    CHECK_RETURNVAL(UNLINK(potato_catcher_handle));

    /* major cleanup */
    CHECK_RETURNVAL(PtlPTFree(ni_logical, logical_pt_index));
    CHECK_RETURNVAL(PtlEQFree(pt_eq_handle));
    CHECK_RETURNVAL(PtlNIFini(ni_logical));
    CHECK_RETURNVAL(libtest_fini());

    PtlFini();

    return 0;
}
コード例 #4
0
ファイル: test_flowctl_noeq.c プロジェクト: foool/portals4
int main(int   argc,
         char *argv[])
{
    ptl_handle_ni_t ni_handle;
    ptl_process_t   *procs;
    int             rank;
    ptl_pt_index_t  pt_index, signal_pt_index;
    HANDLE_T        value_e_handle, signal_e_handle;
    int             num_procs;
    ptl_handle_eq_t eq_handle;
    ptl_handle_ct_t ct_handle;
    ptl_handle_md_t md_handle;

    CHECK_RETURNVAL(PtlInit());

    CHECK_RETURNVAL(libtest_init());

    rank = libtest_get_rank();
    num_procs = libtest_get_size();
    if (num_procs < 2) {
        fprintf(stderr, "test_flowctl_noeq requires at least two processes\n");
        return 77;
    }

    CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
                              PTL_PID_ANY, NULL, NULL, &ni_handle));
    procs = libtest_get_mapping(ni_handle);
    CHECK_RETURNVAL(PtlSetMap(ni_handle, num_procs, procs));


    if (0 == rank) {
        ENTRY_T         value_e;

        /* create data ME */
        CHECK_RETURNVAL(PtlEQAlloc(ni_handle, (num_procs - 1) * ITERS / 2, &eq_handle));
        CHECK_RETURNVAL(PtlPTAlloc(ni_handle, PTL_PT_FLOWCTRL, eq_handle, 5,
                                   &pt_index));
        value_e.start = NULL;
        value_e.length = 0;
        value_e.ct_handle = PTL_CT_NONE;
        value_e.uid = PTL_UID_ANY;
        value_e.options = OPTIONS;
#if INTERFACE == 1
        value_e.match_id.rank = PTL_RANK_ANY;
        value_e.match_bits = 0;
        value_e.ignore_bits = 0;
#endif
        CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_PRIORITY_LIST, NULL, &value_e_handle));

        /* create signal ME */
        CHECK_RETURNVAL(PtlCTAlloc(ni_handle, &ct_handle));
        CHECK_RETURNVAL(PtlPTAlloc(ni_handle, 0, PTL_EQ_NONE, 6,
                                   &signal_pt_index));
        value_e.start = NULL;
        value_e.length = 0;
        value_e.ct_handle = ct_handle;
        value_e.uid = PTL_UID_ANY;
        value_e.options = OPTIONS | PTL_LE_EVENT_SUCCESS_DISABLE | PTL_LE_EVENT_CT_COMM;
#if INTERFACE == 1
        value_e.match_id.rank = PTL_RANK_ANY;
        value_e.match_bits = 0;
        value_e.ignore_bits = 0;
#endif
        CHECK_RETURNVAL(APPEND(ni_handle, 6, &value_e, PTL_PRIORITY_LIST, NULL, &signal_e_handle));
    } else {
        ptl_md_t        md;

        /* 16 extra just in case... */
        CHECK_RETURNVAL(PtlEQAlloc(ni_handle, ITERS * 2 + 16, &eq_handle));

        md.start = NULL;
        md.length = 0;
        md.options = 0;
        md.eq_handle = eq_handle;
        md.ct_handle = PTL_CT_NONE;

        CHECK_RETURNVAL(PtlMDBind(ni_handle, &md, &md_handle));
    }

    libtest_barrier();

    if (0 == rank) {
        ptl_ct_event_t  ct;
        ptl_event_t ev;
        int ret, count = 0, saw_dropped = 0, saw_flowctl = 0;

        /* wait for signal counts */
        CHECK_RETURNVAL(PtlCTWait(ct_handle, num_procs - 1, &ct));
        if (ct.success != num_procs - 1 || ct.failure != 0) {
            return 1;
        }

        /* wait for event entries */
        while (count < ITERS * (num_procs - 1)) {
            ret = PtlEQWait(eq_handle, &ev);
            if (PTL_OK == ret) {
                ;
            } else if (PTL_EQ_DROPPED == ret) {
                saw_dropped++;
                if (ev.type == PTL_EVENT_PT_DISABLED){
                    saw_flowctl++;
                    CHECK_RETURNVAL(PtlPTEnable(ni_handle, pt_index));
                }
                break;
            } else {
                fprintf(stderr, "0: Unexpected return code from EQWait: %d\n", ret);
                return 1;
            }

            if (ev.type == PTL_EVENT_PT_DISABLED) {
                CHECK_RETURNVAL(PtlPTEnable(ni_handle, pt_index));
                saw_flowctl++;
            } else {
                count++;
            }
        }

        fprintf(stderr, "0: Saw %d dropped, %d flowctl\n", saw_dropped, saw_flowctl);
        if (saw_flowctl == 0) {
            return 1;
        }
    } else {
        ptl_process_t target;
        ptl_event_t ev;
        int ret, count = 0, fails = 0;
        int i;
        int *fail_seen;

        fail_seen = malloc(sizeof(int) * ITERS);
        if (NULL == fail_seen) {
             fprintf(stderr, "%d: malloc failed\n", rank);
             return 1;
        }
        memset(fail_seen, 0, sizeof(int) * ITERS);

        target.rank = 0;
        for (i = 0 ; i < ITERS ; ++i) {
            CHECK_RETURNVAL(PtlPut(md_handle,
                                   0,
                                   0,
                                   PTL_ACK_REQ,
                                   target,
                                   5,
                                   0,
                                   0,
                                   (void*)(size_t)i,
                                   0));
            usleep(100);
        }

        while (count < ITERS) {
            ret = PtlEQGet(eq_handle, &ev);
            if (PTL_EQ_EMPTY == ret) {
                continue;
            } else if (PTL_OK != ret) {
                fprintf(stderr, "%d: PtlEQGet returned %d\n", rank, ret);
                return 1;
            }

            if (ev.ni_fail_type == PTL_NI_OK) {
                if (ev.type == PTL_EVENT_SEND) {
                    continue;
                } else if (ev.type == PTL_EVENT_ACK) {
                    count++;
                } else {
                    fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
                }
            } else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
                int iter = (size_t) ev.user_ptr;
                if (fail_seen[iter]++ > 0) {
                    fprintf(stderr, "%d: Double report of PT_DISABLED for "
                            "iteration %d\n", rank, iter);
                    return 1;
                }
                count++;
                fails++;
            } else {
                fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
                return 1;
            }
        }

        fprintf(stderr, "%d: Saw %d of %d events as fails\n", rank, fails, count);

        CHECK_RETURNVAL(PtlPut(md_handle,
                               0,
                               0,
                               PTL_NO_ACK_REQ,
                               target,
                               6,
                               0,
                               0,
                               NULL,
                               0));
        /* wait for the send event on the last put */
        CHECK_RETURNVAL(PtlEQWait(eq_handle, &ev));

        while (fails > 0) {
            CHECK_RETURNVAL(PtlPut(md_handle,
                                   0,
                                   0,
                                   PTL_ACK_REQ,
                                   target,
                                   5,
                                   0,
                                   0,
                                   NULL,
                                   0));
            while (1) {
                ret = PtlEQWait(eq_handle, &ev);
                if (PTL_OK != ret) {
                    fprintf(stderr, "%d: PtlEQWait returned %d\n", rank, ret);
                    return 1;
                }

                if (ev.ni_fail_type == PTL_NI_OK) {
                    if (ev.type == PTL_EVENT_SEND) {
                        continue;
                    } else if (ev.type == PTL_EVENT_ACK) {
                        fails--;
                        break;
                    } else {
                        fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
                    }
                } else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
                    break;
                } else {
                    fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
                    return 1;
                }
            }
        }
    }

    libtest_barrier();

    if (0 == rank) {
        CHECK_RETURNVAL(UNLINK(signal_e_handle));
        CHECK_RETURNVAL(PtlPTFree(ni_handle, signal_pt_index));
        CHECK_RETURNVAL(PtlCTFree(ct_handle));
        CHECK_RETURNVAL(UNLINK(value_e_handle));
        CHECK_RETURNVAL(PtlPTFree(ni_handle, pt_index));
        CHECK_RETURNVAL(PtlEQFree(eq_handle));
    } else {
        CHECK_RETURNVAL(PtlMDRelease(md_handle));
        CHECK_RETURNVAL(PtlEQFree(eq_handle));
    }

    CHECK_RETURNVAL(PtlNIFini(ni_handle));
    CHECK_RETURNVAL(libtest_fini());
    PtlFini();

    return 0;
}
コード例 #5
0
ファイル: test_prepostME.c プロジェクト: foool/portals4
void test_prepostME(int             cache_size,
                    int            *cache_buf,
                    ptl_handle_ni_t ni,
                    int             npeers,
                    int             nmsgs,
                    int             nbytes,
                    int             niters)
{
    int    i, j, k;
    double tmp, total = 0;

    ptl_handle_md_t send_md_handle;
    ptl_md_t        send_md;
    ptl_process_t   dest;
    ptl_size_t      offset;
    ptl_pt_index_t  index;
    ptl_handle_eq_t recv_eq_handle;
    ptl_handle_me_t me_handles[npeers * nmsgs];
    ptl_event_t     event;

    ptl_assert(PtlEQAlloc(ni, nmsgs * npeers + 1,
                          &send_md.eq_handle), PTL_OK);

    send_md.start     = send_buf;
    send_md.length    = SEND_BUF_SIZE;
    send_md.options   = PTL_MD_UNORDERED;
    send_md.ct_handle = PTL_CT_NONE;

    ptl_assert(PtlMDBind(ni, &send_md, &send_md_handle), PTL_OK);

    ptl_assert(PtlEQAlloc(ni, nmsgs * npeers + 1, &recv_eq_handle), PTL_OK);

    ptl_assert(PtlPTAlloc(ni, 0, recv_eq_handle, TestSameDirectionIndex,
                          &index), PTL_OK);

    ptl_assert(TestSameDirectionIndex, index);

    tmp = timer();
    for (j = 0; j < npeers; ++j) {
        for (k = 0; k < nmsgs; ++k) {
            ptl_process_t src;
            src.rank = recv_peers[j];
            postME(ni, index, recv_buf + (nbytes * (k + j * nmsgs)),
                   nbytes, src, magic_tag, &me_handles[k + j * nmsgs]);
        }
    }
    total += (timer() - tmp);

    for (i = 0; i < niters - 1; ++i) {
        cache_invalidate(cache_size, cache_buf);

        libtest_Barrier();

        tmp = timer();
        for (j = 0; j < npeers; ++j) {
            for (k = 0; k < nmsgs; ++k) {
                offset    = (nbytes * (k + j * nmsgs));
                dest.rank = send_peers[npeers - j - 1],
                ptl_assert(libtest_Put_offset(send_md_handle, offset, nbytes,
                                           dest, index, magic_tag, offset), PTL_OK);
            }
        }

        /* wait for sends */
        for (j = 0; j < npeers * nmsgs; ++j) {
            ptl_assert(PtlEQWait(send_md.eq_handle, &event), PTL_OK);
            ptl_assert(event.type, PTL_EVENT_SEND);
        }

        /* wait for receives */
        for (j = 0; j < npeers * nmsgs; j++) {
            PtlEQWait(recv_eq_handle, &event);
        }

        for (j = 0; j < npeers; ++j) {
            for (k = 0; k < nmsgs; ++k) {
                ptl_process_t src;
                src.rank = recv_peers[j];
                postME(ni, index, recv_buf + (nbytes * (k + j * nmsgs)),
                       nbytes, src, magic_tag, &me_handles[k + j * nmsgs]);
            }
        }
        total += (timer() - tmp);
    }

    libtest_Barrier();

    tmp = timer();
    for (j = 0; j < npeers; ++j) {
        for (k = 0; k < nmsgs; ++k) {
            offset    = (nbytes * (k + j * nmsgs));
            dest.rank = send_peers[npeers - j - 1],
            ptl_assert(libtest_Put_offset(send_md_handle, offset, nbytes, dest,
                                       index, magic_tag, offset), PTL_OK);
        }
    }
    /* wait for sends */
    for (j = 0; j < npeers * nmsgs; ++j) {
        ptl_assert(PtlEQWait(send_md.eq_handle, &event), PTL_OK);
        ptl_assert(event.type, PTL_EVENT_SEND);
    }

    /* wait for receives */
    for (j = 0; j < npeers * nmsgs; j++) {
        PtlEQWait(recv_eq_handle, &event);
    }

    total += (timer() - tmp);

    ptl_assert(PtlEQFree(send_md.eq_handle), PTL_OK);

    ptl_assert(PtlMDRelease(send_md_handle), PTL_OK);

    ptl_assert(PtlEQFree(recv_eq_handle), PTL_OK);

    ptl_assert(PtlPTFree(ni, index), PTL_OK);

    tmp = libtest_AllreduceDouble(total, PTL_SUM);
    display_result("pre-post", (niters * npeers * nmsgs * 2) / (tmp / world_size));
}
コード例 #6
0
ファイル: test_unlink.c プロジェクト: foool/portals4
int main(int   argc,
         char *argv[])
{
    ptl_handle_ni_t ni_h;
    ptl_pt_index_t  pt_index;
    uint64_t       *buf;
    ENTRY_T         entry;
    HANDLE_T        entry_h;
    ptl_md_t        md;
    ptl_handle_md_t md_h;
    int             rank;
    int             num_procs;
    int             ret;
    ptl_process_t  *procs;
    ptl_handle_eq_t eq_h;
    ptl_event_t     ev;
    ptl_hdr_data_t rcvd = 0;
    ptl_hdr_data_t goal = 0;
    ptl_hdr_data_t hdr_data = 1;
    ptl_size_t offset = sizeof(uint64_t);
    uint32_t distance;
    int sends = 0;

    CHECK_RETURNVAL(PtlInit());
    CHECK_RETURNVAL(libtest_init());

    rank = libtest_get_rank();
    num_procs = libtest_get_size();

    /* This test only succeeds if we have more than one rank */
    if (num_procs < 2) return 77;

    CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
                              PTL_PID_ANY, NULL, NULL, &ni_h));

    procs = libtest_get_mapping(ni_h);
    CHECK_RETURNVAL(PtlSetMap(ni_h, num_procs, procs));

    CHECK_RETURNVAL(PtlEQAlloc(ni_h, 1024, &eq_h));
    CHECK_RETURNVAL(PtlPTAlloc(ni_h, 0, eq_h, 0, &pt_index));
    assert(pt_index == 0);
    
    buf = malloc(sizeof(uint64_t) * num_procs);
    assert(NULL != buf);

    md.start = buf;
    md.length = sizeof(uint64_t) * num_procs;
    md.options = PTL_MD_UNORDERED;
    md.eq_handle = eq_h;
    md.ct_handle = PTL_CT_NONE;
    CHECK_RETURNVAL(PtlMDBind(ni_h, &md, &md_h));

    entry.start = buf;
    entry.length = sizeof(uint64_t) * num_procs;
    entry.ct_handle = PTL_CT_NONE;
    entry.uid = PTL_UID_ANY;
    entry.options = OPTIONS;
#if MATCHING == 1
    entry.match_id.rank = PTL_RANK_ANY;
    entry.match_bits = 0;
    entry.ignore_bits = 0;
    entry.min_free = 0;
#endif
    CHECK_RETURNVAL(APPEND(ni_h, pt_index, &entry,
                           PTL_PRIORITY_LIST, NULL, &entry_h));

    /* ensure ME is linked before the barrier */
    CHECK_RETURNVAL(PtlEQWait(eq_h, &ev));
    assert( ev.type == PTL_EVENT_LINK );

    libtest_barrier();

    /* Bruck's Concatenation Algorithm */
    memcpy(buf, &rank, sizeof(uint64_t));
    for (distance = 1; distance < num_procs; distance *= 2) {
        ptl_size_t to_xfer;
        int peer;
        ptl_process_t proc;

        if (rank >= distance) {
            peer = rank - distance;
        } else {
            peer = rank + (num_procs - distance);
        }

        to_xfer = sizeof(uint64_t) * MIN(distance, num_procs - distance);
        proc.rank = peer;
        CHECK_RETURNVAL(PtlPut(md_h, 
                               0, 
                               to_xfer, 
                               PTL_NO_ACK_REQ, 
                               proc,
                               0,
                               0,
                               offset,
                               NULL,
                               hdr_data));
        sends += 1;

        /* wait for completion of the proper receive, and keep count
           of uncompleted sends.  "rcvd" is an accumulator to deal
           with out-of-order receives, which are IDed by the
           hdr_data */
        goal |= hdr_data;
        while ((rcvd & goal) != goal) {
            ret = PtlEQWait(eq_h, &ev);
            switch (ret) {
            case PTL_OK:
                if (ev.type == PTL_EVENT_SEND) {
                    sends -= 1;
                } else {
                    rcvd |= ev.hdr_data;
                    assert(ev.type == PTL_EVENT_PUT);
                    assert(ev.rlength == ev.mlength);
                    assert((ev.rlength == to_xfer) || (ev.hdr_data != hdr_data));
                }
                break;
            default:
                fprintf(stderr, "PtlEQWait failure: %d\n", ret);
                abort();
            }
        }
        
        hdr_data <<= 1;
        offset += to_xfer;
    }

    /* wait for any SEND_END events not yet seen */
    while (sends) {
        ret = PtlEQWait(eq_h, &ev);
        switch (ret) {
        case PTL_OK:
            assert( ev.type == PTL_EVENT_SEND );
            sends -= 1;
            break;
        default:
            fprintf(stderr, "PtlEQWait failure: %d\n", ret);
            abort();
        }
    }

    CHECK_RETURNVAL(UNLINK(entry_h));
    CHECK_RETURNVAL(PtlMDRelease(md_h));
    free(buf);

    libtest_barrier();

    /* cleanup */
    CHECK_RETURNVAL(PtlPTFree(ni_h, pt_index));
    CHECK_RETURNVAL(PtlEQFree(eq_h));
    CHECK_RETURNVAL(PtlNIFini(ni_h));
    CHECK_RETURNVAL(libtest_fini());
    PtlFini();

    return 0;
}