CAKE
Projects | | Links: Tech Report | Source
Surved on all major Scheduling Frameworks, such as Borg, YARN, Paragon, Omega, Mesos, etc. Proposed another frameworkd using AI(Genetic Algortihm) and CUDA power.
A good plan consumes all available facts during decision making. An excellent plan can be intelligent if it takes further inferences from available facts. In short, we are talking about Artificial Intelligence in Scheduling. We see AI is the new electricity for any application. Scheduling is the key to an Application and Organization Success. A good scheduling will make the application faster if and only if the scheduling is fast. Organization benefits if scheduling maximizes the cluster utilization. In short, scheduling’s key idea is to maximize cluster utilization use and minimize latency. The existing framework does proper scheduling but uses the only tiny portion of Knowledge.
Cluster Allocation for Kernel Execution This paper summarizes our study of many scheduler architectures, and our own design of new pluggable scheduler framework called CAKE, which uses Genetic Algorithm, an Artificial Intelligence technique, for scheduling. We have presented our thoughts about why we need CAKE framework and we have also mentioned few self-critics for it.
We have also given proof of concept for CAKE using Genetic Algorithm. Our research notes section describes why advanced concepts (AI, etc) in scheduling must be unified by all researchers in Apache Open Source project, such that any application or organization will not worry about scheduling at all.