Multi-GPU server

This is a minimal working example of a multi-threaded, multi-GPU Waffle web server.

Waffle is a fast, asynchronous, express-inspired web framework for Lua/Torch built on top of ASyNC: https://github.com/benglard/waffle .

Setup

You must have the following up-to-date packages installed: torch, nn, cutorch, cunn, paths, threads, async, waffle, graphicsmagick

Torch is here: http://torch.ch/docs/getting-started.html#installing-torch .

After Torch is installed, most of the other packages can be installed in the following manner:

	luarocks install cutorch

Waffle (and its depedency htmlua) can be found here: https://github.com/benglard/waffle#installation .

Before using luarocks to install the GraphicsMagick wrapper, you must have GraphicsMagick installed. On an Ubuntu/Debian system, this is simply:

	sudo apt-get install graphicsmagick

On Mac OSX:

	brew install graphicsmagick

Use

To turn on the webserver:

	th multithread_server.lua

Testing/Profiling

The script called multithread_posts.sh provides a test of multithread_server.lua by sending large numbers of POST requests in parallel. For example, to send 500 POST requests in batches of 50 parallel requests:

	./multithread_posts.sh 50 500 random_image.jpg

To create a CPU/GPU timeline, first create an executable bash script called start_server.sh to run the server:

	#!/bin/bash
	/path/to/th /path/to/multithread_server.lua

Then, from the command line, run the following:

	nvprof --profile-child-processes -o timeline_%p /path/to/start_server.sh

This will create the timeline. To view it, start up nvvp and import the timeline.