Teams of 2 are ok.

Use the ideas we've seen to implement a parallel text histogram program, both on the multicore CPU and on the manycore GPU. Do it with and without private subtotal arrays.

Use current tools, like managed memory on the device.

Compare the performance processing a big file. Separate the time to read the data into memory from the time to process it.

Try to use debugging tools to find hotspots.

To reduce file reading time, before starting your program, put the data file into the in-core filesystem in /dev/shm/, aka /ds .