hcho3/skiplist_cuda

A parallel (CUDA) implementation of skiplist

skiplist_cuda: A parallel/CUDA implementation of skiplist

Skiplists are a variant of linked lists that allow insertions in O(log n) time. Inspired by a GTC 2013 talk, we build a parallel implementation of skiplist where multiple GPU threads can insert simultaneously. We make heavy use of atomic compare-and-swap operation. By default, our skiplist contains integers; to use different types, modify the type definition E in skip_parallel.h.

We assume that you are using a 64-bit machine.

See here for a theoretical analysis of skiplists.

Usage

For more complete usage, see tester_parallel.cu.

Create a skiplist on the host:

#include "spmat_parallel.h"
...
// set heap size of 128 MB
cudaDeviceSetLimit(cudaLimitMallocHeapSize, 128*1024*1024);
Skiplist *sl = skiplist_create();
...

Important: It is crucial to set heap size large enough to contain all the elements to be inserted. When skiplist_insert() crashes in the middle, try increasing the heap size.

Insert elements to the skiplist Simply call skiplist_insert(). The following code stub will insert all elements in the array a into the skiplist sl:

#include "spmat_parallel.h"
...
__global__ void add(Skiplist *sl, int *a, int N)
{
  int x = threadIdx.x + blockIdx.x * blockDim.x;

  while (x < N) {
    skiplist_insert(sl, a[x]);
    x += blockDim.x * gridDim.x;
  }
}
...
add<<<100, 320>>>(sl, a_dev, N);

Remove elements from the list: Unfortunately, this feature is not completed yet. We'll try to make some time in the future.

Collect the content to the host:

E *result = skiplist_gather(sl, &result_dim);
// result points to a host buffer
// result_dim contains the number of elements

This function will allocate a buffer on the host and return its pointer. At the same time, it will return the number of elements via the second argument.

De-allocate the skiplist

skiplist_destroy(sl);

How to compile the library

make

This will compile the library, along with a sample program in tester_parallel.cu.

How to link your program with the library

nvcc -o [your executable] [your object files] skip_parallel.o safety.o

safety.o contains a little macro that checks the return value of all CUDA API functions.

Credits

Hyunsu "Philip" Cho and Sam Johnson, Trinity College, Hartford, CT