Software Systems

As an end-of-the-year treat, we present to the SRE community some of the more beautiful images we have seen in our monitoring system. These images offer a glimpse into the visual patterns that appear in our variables and time-series, and the beauty that emerges from chaos (Truly: some of these images appeared during difficult rollouts, or even during incidents).

Power consumption costs of computer fleets can be one of the main operational costs of medium or large-sized office sites. In order to optimise the power consumption of the fleet, a set of optimal power management policies must be generated and enforced. Generating these policies is an optimisation problem of finding the power off timeout value for each computer that maximises energy savings...

Accurate hardware performance models are critical to efficient code generation. They can be used by compilers to make heuristic decisions, by superoptimizers as an minimization objective, or by autotuners to find an optimal configuration of a specific program. However, they are difficult to develop because contemporary processors are complex, and the recent proliferation of deep learning...

At Google, we would like to develop a cost model that can accurately estimate the execution time of a machine learning model running on a Tensor Processing Unit (TPU). This cost model can be used by a compiler to make heuristic decisions, by an autotuner to find an optimal configuration of a specific program, and by Neural Architecture Search to co-optimize accuracy and inference time. However,...