GPUs (cuda)

See examples at https://github.com/hpc-cofc/example-runs/tree/master/01_prime-number-finder-on-CPUs-and-GPUs

Prime number finder

Idea

This calculation finds and counts the number of prime numbers below 500,000 using

• Sieve of Eratosthenes method implemented by Boris Bershadsky

• and an MPI-parallelized brute-force method implemented by John Burkardt@FSU

Tests

In short, we run 1. Sieve of Eratosthenes - GPUs - CUDA Multithreading version 2. Sieve of Eratosthenes - CPUs - simple CPU version 3. Brute-force method - CPUs - MPI-parallelized version

How to run

You can use the test-queue.sh script to run the calculations on different queues/partitions on the CofC HPC cluster by ./test-queue.sh

Results

Here is a summary from a sample run.

Sieve of Eratosthenes - GPUs vs CPUs

See summary-gpuq.dat

Calculation	time (milliseconds) [lower is better]
GPU run	37
CPU run	278

Brute-force method - CPUs - MPI-parallelized version

See summary-stdmemq.dat

Number of Cores or MPI processes	Real time (seconds) [lower is better]	Parallel efficiency (higher is better)
2	33	0.97
4	16	0.80
8	8	0.93
16	5	0.88