Moscow State University Scientists Speed Up Training of Computing Management Systems
Researchers at the Faculty of Computational Mathematics and Cybernetics of Lomonosov Moscow State University have developed approaches that reduce the time required to train systems that manage the distribution of computing tasks across a network.
The research focuses on scheduling computing tasks under changing workloads, including in networks that process data from Internet of Things devices. In such networks, tasks arrive unpredictably, requiring workloads to be distributed dynamically across computing nodes. Researchers had previously proposed a multi-agent reinforcement learning approach, but training such systems requires substantial computing resources: a single training run can take about 10 hours of processor time, and parameter tuning requires many such runs.
The researchers identified the main sources of computational overhead and proposed solutions, including eliminating unnecessary data copying and running computations in parallel.
The time required to configure the system was reduced to several days. The research was presented at the Lomonosov Readings conference.
