Kernel launches blocking when 1024 kernels in a queue

Hi all,

our application processes large data with a use of many small kernels. We process the data in bursts because some kernels take a performance advantage when running over a whole burst (but some others don’t - the number of them is relative to a burst size). This means the larger the burst can be the more performance we can get.

But we found that we cannot increase the burst size as much as we would sometimes like. That’s because when many kernels are waiting in a queue, the new kernel launches become blocking. This is what we would definitely need to avoid because while the kernels run we need to prepare other bursts on a cpu.

It’s not possible to find the “ok” burst size as we actually don’t know the number of kernels in a queue and also the device’s performance vary so for different devices a different number of kernels would be finished in the same code point.

Now we found the actual number of kernels in a queue when the others become blocking, is 1024. You can try with a simple example (slightly modified vectorAdd):

/**
 * Copyright 1993-2013 NVIDIA Corporation.  All rights reserved.
 *
 * Please refer to the NVIDIA end user license agreement (EULA) associated
 * with this source code for terms and conditions that govern your use of
 * this software. Any use, reproduction, disclosure, or distribution of
 * this software and related documentation outside the terms of the EULA
 * is strictly prohibited.
 *
 */

/**
 * Vector addition: C = A + B.
 *
 * This sample is a very basic sample that implements element by element
 * vector addition. It is the same as the sample illustrating Chapter 2
 * of the programming guide with some additions like error checking.
 */

#include <stdio.h>

// For the CUDA runtime routines (prefixed with "cuda_")
#include <cuda_runtime.h>

/**
 * CUDA Kernel Device code
 *
 * Computes the vector addition of A and B into C. The 3 vectors have the same
 * number of elements numElements.
 */
__global__ void
vectorAdd(const float *A, const float *B, float *C, int numElements)
{
    int i = blockDim.x * blockIdx.x + threadIdx.x;

    if (i < numElements)
    {
        C[i] = A[i] + B[i];
    }
}

/**
 * Host main routine
 */
int
main(void)
{
    // Error code to check return values for CUDA calls
    cudaError_t err = cudaSuccess;

    // Print the vector length to be used, and compute its size
    int numElements = 50000000; // make the size large enough for the kernels to last some time (to overcome the cpu kernel launch overhead ~ 7us)
    size_t size = numElements * sizeof(float);
    printf("[Vector addition of %d elements]\n", numElements);

    // Allocate the host input vector A
    float *h_A = (float *)malloc(size);

    // Allocate the host input vector B
    float *h_B = (float *)malloc(size);

    // Allocate the host output vector C
    float *h_C = (float *)malloc(size);

    // Verify that allocations succeeded
    if (h_A == NULL || h_B == NULL || h_C == NULL)
    {
        fprintf(stderr, "Failed to allocate host vectors!\n");
        exit(EXIT_FAILURE);
    }

    // Initialize the host input vectors
    for (int i = 0; i < numElements; ++i)
    {
        h_A[i] = rand()/(float)RAND_MAX;
        h_B[i] = rand()/(float)RAND_MAX;
    }

    // Allocate the device input vector A
    float *d_A = NULL;
    err = cudaMalloc((void **)&d_A, size);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to allocate device vector A (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Allocate the device input vector B
    float *d_B = NULL;
    err = cudaMalloc((void **)&d_B, size);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to allocate device vector B (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Allocate the device output vector C
    float *d_C = NULL;
    err = cudaMalloc((void **)&d_C, size);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to allocate device vector C (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Copy the host input vectors A and B in host memory to the device input vectors in
    // device memory
    printf("Copy input data from the host memory to the CUDA device\n");
    err = cudaMemcpy(d_A, h_A, size, cudaMemcpyHostToDevice);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to copy vector A from host to device (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    err = cudaMemcpy(d_B, h_B, size, cudaMemcpyHostToDevice);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to copy vector B from host to device (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }
    
    // launch 1030 kernels
    for (int i = 0; i < 1030; i++) {
        // Launch the Vector Add CUDA Kernel
        int threadsPerBlock = 256;
        int blocksPerGrid =(numElements + threadsPerBlock - 1) / threadsPerBlock;
//         printf("CUDA kernel launch with %d blocks of %d threads\n", blocksPerGrid, threadsPerBlock);
        vectorAdd<<<blocksPerGrid, threadsPerBlock>>>(d_A, d_B, d_C, numElements);
    }
    
    err = cudaGetLastError();
    
    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to launch vectorAdd kernel (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Copy the device result vector in device memory to the host result vector
    // in host memory.
    printf("Copy output data from the CUDA device to the host memory\n");
    err = cudaMemcpy(h_C, d_C, size, cudaMemcpyDeviceToHost);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to copy vector C from device to host (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Verify that the result vector is correct
    for (int i = 0; i < numElements; ++i)
    {
        if (fabs(h_A[i] + h_B[i] - h_C[i]) > 1e-5)
        {
            fprintf(stderr, "Result verification failed at element %d!\n", i);
            exit(EXIT_FAILURE);
        }
    }
    printf("Test PASSED\n");

    // Free device global memory
    err = cudaFree(d_A);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to free device vector A (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }
    err = cudaFree(d_B);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to free device vector B (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }
    err = cudaFree(d_C);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to free device vector C (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Free host memory
    free(h_A);
    free(h_B);
    free(h_C);

    // Reset the device and exit
    err = cudaDeviceReset();

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to deinitialize the device! error=%s\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    printf("Done\n");
    return 0;
}

Here’s what profiler will show:
External Media

So our questions are:

  • Does anyone please know how to workaroud this behavior? We tried to split the kernels to streams but the behavior is still the same.
  • Is it possible somehow to get the number of kernels waiting in a queue so we could detect easily the point where it would block?
  • Is this problem at least documented somewhere? We tried but haven't found it anywhere.
  • Thanks a lot.

    If you can’t increase the size of your kernels or launch larger grids, then you could marshal your various kernels – func pointer + args + grid dimensions – into a ring buffer that is serviced by a separate blocking thread.

    This is duplicating what the driver is already doing for you but it would let you create an arbitrarily large “user stream” at the cost of adding a lot of complexity to your application.

    Probably not a good idea unless you have an especially unique use case. :)

    Yes, you can pepper your stream(s) with events and query for completion with cudaEventQuery().

    Hi allanmac, thanks for your answer.

    We’ve been thinking about that already but it means major changes to our code and therefore probably many new issues. We tried to use dynamic parallelism for this as it seems the number of kernels launched from another kernel doesn’t count to that queue so we may be possible to launch more kernels than now without hitting the limit. But here we found another problem: Dynamically launched kernels ~ 2.5x slower

    Great, that may be possible. Thanks for a suggestion. But again it involves quite many changes as we use to call varying number of kernels from many places in the code.

    Am I right with the limit of 1024 launches? To detect the bad behavior the exact number is crucial. We would really appreciate some documentation on this.

    Thanks again for your suggestions!

    The limit might depend on device and CUDA driver version. I recall in the past (several software and hardware revisions ago) reading that someone empirically discovered the limit was 32 or 64.