This is the submission of Luca Venturi for the GPU Specialization Capstone Project

My company developed a custom, columnar NoSql DB for analytics; this program is intended to test some fo the basic functionalities of a similar DB, and evaluate the possible speed-up of using a GPU.

The program define an abstract class, DBEngine, that defines some expected functions, without implementing them. There is a class called DBEngineCPU, that implements it with a single thread, both as a speed reference and to verify the correctness of the other solutions. DBEngineCPU4Threads implements a DBEngine using 4 CPU threads. The GPU implementation using nVidia NPP is provided by DBEngineNPP. In addition, there is a class called DBEngineNPPInit, which is used to initialize the NPP libraries.

The program has been developed and tested on WSL, using Ubuntu 18.04.

To build it, you can use the command make build.

To run it, you could do it in different ways, for example:

• ./run.sh
• make run
• ./luca

The following is the output after creating 1 000 000 000 of random numbers: CPU - Time: 12794 ms - Init: 0 ms - Copy: 0 - Count: 4653 - Average: -11448.7 - Min: -2147483648 - Max: 2147483643 CPU 4 - Time: 4052 ms - Init: 0 ms - Copy: 0 - Count: 4653 - Average: -11448.7 - Min: -2147483648 - Max: 2147483643 NPP INIT - Time: 1136 ms - Init: 472 ms - Copy: 544 - Count: 4653 - Average: -11449 - Min: -2147483648 - Max: 2147483643 NPP - Time: 849 ms - Init: 146 ms - Copy: 558 - Count: 4653 - Average: -11449 - Min: -2147483648 - Max: 2147483643 NPP - pinned - Time: 4382 ms - Init: 1353 ms - Copy: 357 - Count: 4653 - Average: -11449 - Min: -2147483648 - Max: 2147483643

This is with 100 000 000 random numbers: CPU - Time: 1263 ms - Init: 0 ms - Copy: 0 - Count: 469 - Average: -19081.6 - Min: -2147483627 - Max: 2147483615 CPU 4 - Time: 457 ms - Init: 0 ms - Copy: 0 - Count: 469 - Average: -19081.6 - Min: -2147483627 - Max: 2147483615 NPP INIT - Time: 533 ms - Init: 432 ms - Copy: 55 - Count: 469 - Average: -19082 - Min: -2147483627 - Max: 2147483615 NPP - Time: 75 ms - Init: 2 ms - Copy: 65 - Count: 469 - Average: -19082 - Min: -2147483627 - Max: 2147483615 NPP - pinned - Time: 388 ms - Init: 118 ms - Copy: 34 - Count: 469 - Average: -19082 - Min: -2147483627 - Max: 2147483615

• The GPU can speed-up these tasks
• NPP can be very valuable for these tasks, and it is much easier than trying to do it with the Runtime API
• Pinned memory has faster transfers, but slower allocation
• Pinned memory has some issue (maybe related to size and access patterns) that potentially make it quite tricky to use
• GPU code is much longer and low level than CPU code