Low-overhead decision support for dynamic buffer reallocation

Abstract Effective I/O buffering is a performance-critical task in database management systems. Accordingly, systems usually employ various special-purpose buffers to align, e.g., device speed, page size, and replacement policies with the actual data and workload. However, such partitioning of available buffer memory results in complex optimiza-tion problems for database administrators and also in fragile configurations which quickly deteriorate on workload shifts. Reliable forecasts of I/O costs enable a system to evaluate al-ternative configurations to continuously optimize its buffer memory allocation at runtime. So far, all techniques pro-posed for the prediction of buffer performance focus solely on hit ratio gains for increased buffer sizes to identify buffers which promise the greatest benefit. These approaches, how-ever, assume that their forecast allows to extrapolate the ef-fect for buffer downsizing, too. As we will show, this comes along with a severe risk of wrong tuning decisions, which may heavily impact system performance. Thus, we empha-size the importance of reliably forecasting the penalty to expect for shrinking buffers in favor of others. We explore the use of lightweight extensions for widely used buffer algorithms to perform on-the-fly simulation of buffer per-formance of smaller and larger buffer sizes simultaneously. Furthermore, we present a simple cost model and demon-strate how to compose these concepts into a self-tuning com-ponent for dynamic buffer reallocation. This is an extended version of the paper “Lightweight Performance Forecasts for Buffer Algorithms ” selected for the special CSRD issu