performance_engineering

The code in this repo is a part of performance engineering presentation I gave at Celantur.

• Basic MP - basic setup for multiprocessing in Python. Consists of a worker object that is connected to main process via multiprocessing queue.
• Cache miss - demonstration of substantial slowdown of operations when we have a lot of cache misses. Used to demosntrate how slow the exchange works between CPU and RAM
• Context switch - Just a program that runs multiple threads in parallel. Shows that the performance drops after N_threads > N_cores
• Subprocess manipulation - serves as an example that OS can send signals to subprocesses, as opposed to threads
• thrust.cu - Shows the easy way to use GPU as processing of vector operations
• FlaskAPI/ApiForApi/SanicApi - show the difference between IO and compute bounded tasks. Displays that compute bound tasks can be improved with the multiprocessing and the IO bound task with asyncronous processing.
• intro.cu - basic example program that runs on GPU. Does memory allocation, host-to-device copy, some simple kernel, device-to-host copy and memory free. Used to display the necessety of profiling (done with Nsight)
• async_mem_cpy_streams.cu - works on top of intro.cu. Shows the possible optimisations for processing, such as simultaneous kernel and memory transfers, staircase optimisation pattern, keeping allocated memory, demostrating the GPU streams.
• nsight.py, nsight_mp.py, nsight_no_nvtx.py - serve as the example of using NSight profiler for torch programs and nesessity of NVTX decorators. Shows that NSight profiling can work with multhreaded and multiprocessed programs.