为什么两个 CUDA Streams 中的操作不重叠？

Question

My program is a pipeline, which contains multiple kernels and memcpys. Each task will go through the same pipeline with different input data. The host code will first chooses a Channel, an encapsulation of scratchpad memory and CUDA objects, when it process a task. And after the last stage, I will record an event then will go to process next task.
The main pipeline logic is in the following. The problem is that operations in different streams are not overlapping. I attached the timeline of processing 10 tasks. You can see none operations in streams are overlapped. For each kernel, there is 256 threads in a block and 5 blocks in a grid. All buffers used for memcpy are pinned, I am sure that I have meet thoserequirements for overlapping kernel execution and data transfers. Can someone help me figure out the reason? Thanks.

Environment information
GPU: Tesla K40m (GK110)
Max Warps/SM: 64
Max Thread Blocks/SM: 16
Max Threads/SM: 2048
CUDA version: 8.0

    void execute_task_pipeline(int stage, MyTask *task, Channel *channel) {
    assert(channel->taken);
    assert(!task->finish());

    GPUParam *para = &channel->para;

    assert(para->col_num > 0);
    assert(para->row_num > 0);

    // copy vid_list to device
    CUDA_ASSERT( cudaMemcpyAsync(para->vid_list_d, task->vid_list.data(),
                sizeof(uint) * para->row_num, cudaMemcpyHostToDevice, channel->stream) );

    k_get_slot_id_list<<<WK_GET_BLOCKS(para->row_num),
        WK_CUDA_NUM_THREADS, 0, channel->stream>>>(
                vertices_d,
                para->vid_list_d,
                para->slot_id_list_d,
                config.num_buckets,
                para->row_num);

    k_get_edge_list<<<WK_GET_BLOCKS(para->row_num),
        WK_CUDA_NUM_THREADS, 0, channel->stream>>>(
                vertices_d,
                para->slot_id_list_d,
                para->edge_size_list_d,
                para->offset_list_d,
                para->row_num);

    k_calc_prefix_sum(para, channel->stream);

    k_update_result_table_k2u<<<WK_GET_BLOCKS(para->row_num),
        WK_CUDA_NUM_THREADS, 0, channel->stream>>>(
            edges_d,
            para->vid_list_d,
            para->updated_result_table_d,
            para->prefix_sum_list_d,
            para->offset_list_d,
            para->col_num,
            para->row_num);

    para->col_num += 1;
    // copy result back to host
    CUDA_ASSERT( cudaMemcpyAsync(&(channel->num_new_rows), para->prefix_sum_list_d + para->row_num - 1,
            sizeof(uint), cudaMemcpyDeviceToHost, channel->stream) );
    // copy result to host memory
    CUDA_ASSERT( cudaMemcpyAsync(channel->h_buf, para->updated_result_table_d,
                channel->num_new_rows * (para->col_num + 1), cudaMemcpyDeviceToHost, channel->stream) );

    // insert a finish event in the end of pipeline
    CUDA_ASSERT( cudaEventRecord(channel->fin_event, channel->stream) );
}

Answer 1

are you trying to overlap treatments which are during 82microsecs ?

Since you have profiled your application, the clue can be in the big orange box between two kernel execution (wich is not readable in your image).

If this is a synchronisation remove it.

If this is a trace like cudaLaunch_KernelName, try to make your treatments bigger (more datas or more computations) because you take more time sending an order to the GPU than it takes to execute it, so you can't make parallel computations into these different streams.