Deploying AI on microcontrollers (MCUs) is challenging. We introduce MIMaaS, a Microcontroller-as-a-Service platform that enables users to upload a model, select a target device, and receive a detailed performance report remotely. A key innovation is our measurement of real-world power consumption, alongside latency and memory usage, directly from the physical hardware. MIMaaS empowers researchers and developers to easily create and validate hardware-aware AI models without needing physical hardware access.

In resource marking theory, users adapt their rates according to their utility functions and congestion signals from the network. Using a special type of utility function, this leads to a proportionally fair rate allocation among the users. We examine the ability of two proposed resource marking implementations, random exponential marking (see Athuraliya, S. et al., Teletraffic Science and Engineering, vol. 4, p.817-28, 2001) and the virtual queue mechanism (see Gibbens, R.J. and Kelly, F.P., Automatica, vol.35, p.1969-85, 1999), to yield a proportionally fair rate allocation. We also propose a third and less complex algorithm, single bit resource marking, that combines the advantages of both.