Improving OFI Transport Counter Polling Implementation

[markbrown314]

The current libfabric counter polling implementation requires a series of separate lock acquisitions fi_cntr(), fi_cntr_err(), and fi_cq_read(), that can be simplified if we only use fi_cnt_wait(). We see improvements in put latency benchmarks when the shmem_transport_ofi_put_quiet() polling code is removed and no change in functionality.

Original:

void shmem_transport_put_quiet(shmem_transport_ctx_t* ctx)
{
    ...
    while (poll_count < shmem_transport_ofi_put_poll_limit ||
           shmem_transport_ofi_put_poll_limit < 0) {
        success = fi_cntr_read(ctx->put_cntr);
        fail = fi_cntr_readerr(ctx->put_cntr);
        cnt = SHMEM_TRANSPORT_OFI_CNTR_READ(&ctx->pending_put_cntr);

        shmem_transport_probe();

        if (success < cnt && fail == 0) {
            SHMEM_TRANSPORT_OFI_CTX_UNLOCK(ctx);
            SPINLOCK_BODY();
            SHMEM_TRANSPORT_OFI_CTX_LOCK(ctx);
        } else if (fail) {
            RAISE_ERROR_MSG("Operations completed in error (%" PRIu64 ")\n", fail);
        } else {
            SHMEM_TRANSPORT_OFI_CTX_UNLOCK(ctx);
            return;
        }
        poll_count++;
    }
    cnt_new = SHMEM_TRANSPORT_OFI_CNTR_READ(&ctx->pending_put_cntr);
    do {
        cnt = cnt_new;
        ssize_t ret = fi_cntr_wait(ctx->put_cntr, cnt, -1);
        cnt_new = SHMEM_TRANSPORT_OFI_CNTR_READ(&ctx->pending_put_cntr);
        OFI_CTX_CHECK_ERROR(ctx, ret);
    } while (cnt < cnt_new);
    shmem_internal_assert(cnt == cnt_new);
    ...

Simplified:

void shmem_transport_put_quiet(shmem_transport_ctx_t* ctx)
{
    ...
    uint64_t cnt, cnt_new;

    cnt_new = SHMEM_TRANSPORT_OFI_CNTR_READ(&ctx->pending_put_cntr);
    do {
        cnt = cnt_new;
        ssize_t ret = fi_cntr_wait(ctx->put_cntr, cnt, -1);
        cnt_new = SHMEM_TRANSPORT_OFI_CNTR_READ(&ctx->pending_put_cntr);
        OFI_CTX_CHECK_ERROR(ctx, ret);
    } while (cnt < cnt_new);
    shmem_internal_assert(cnt == cnt_new);
    ...
}