Daniel Mendoza

The profiliferation of machine learning inference to many domains has highlighted the importance of developing inference serving systems which meet applications' service level objectives. To improve resource utilization, inference serving systems often co-locate multiple models on the same backend machine. However, co-location can cause latency degradation due to interference and can subsequently violate latency requirements. In collaboration with Professor Christos Kozyrakis lab, we characterized the challenges of interference-aware scheduling in inference serving systems and demonstrated its potential benefits.

SystemC is a generic discrete event simulation framework often used for embedded system models with hardware and software components. Although SystemC enables model developers to simulate concurrent processes, implementations of SystemC simulators execute all processes sequentially and thus do not exploit the inherent parallelism of the model to speedup simulation. Thus the Recoding Infrastructure for SystemC (RISC) project was introduced to enable advanced analysis and aggressive parallel simulation of SystemC models. Under mentorship of Professor Rainer Doemer, we developed efficient scheduling algorithms for parallel simulation, integrated parallel SystemC into virtual platform simulation, and developed tools for visualizing SystemC models.