Analyzing the Impact of Concurrency on Scaling Machine Learning Programs Using TensorFlow

In recent times, computer scientists and technology companies have quickly begun to realize that machine learning and creating computer software that is capable of reasoning for itself (at least in theory). What was once only considered science fiction lore is now becoming a reality in front of our very eyes. With this type of computational capability at our disposal, we are left with the question of how best to use it and where to start in creating models that can help us best utilize it. TensorFlow is an open source software library used in machine learning developed and released by Google. It was created by the company in order to help them meet their expanding needs to train systems that can build and detect neural networks for pattern recognition that could be used in their services. It was first released by the Google Brain Team in November 2015 and, at the time of the preparation of this research, the project is still being heavily developed by programmers and researchers both inside of Google and around the world. Thus, it is very possible that some future releases of the software package could remove and/or replace some current capabilities. The point of this thesis is to examine how machine learning programs written with TensorFlow that do not scale well (such as large-scale neural networks) can be made more scalable by using concurrency and distribution of computation among threads. To do this, we will be using lock elision using conditional variables and locking mechanisms (such as semaphores) to allow for smooth distribution of resources to be used by the architecture. We present the trial runs and results of the added implementations and where the results fell short of optimistic expectation. Although TensorFlow is still a work in progress, we will also address where this framework was insufficient in addressing the needs of programmers attempting to write scalable code and whether this type of implementation is sustainable.