Cache eviction policies improve performance by considering object sizes
Caching Policies based on Object Sizes
Performance
Summary
Caches temporarily store data to speed up computer systems, but deciding what to keep or remove can affect performance. This paper looks at methods that consider the size of each piece of data when deciding what to remove from the cache. The authors test these methods using both simple models and real-life data to see how well they work. They also develop new ways to predict the most useful data to keep, improving how caches manage storage
What this means in practice
- •For network schedulers: Manage router caches by using size-aware eviction to improve bandwidth and latency under varied network traffic patterns.
- •For database administrators: Optimize in-memory database caching by applying length-aware policies to better handle mixed-size data workloads.
Authors
Ridwanul Hasan Tanvir, George Kesidis
Abstract
We evaluate eviction policies that consider the lengths of the objects being cached for different cache sizes (cost) and different workloads. Workloads include the Independent Reference Model and real-world production traces. In particular, hybrid policies are considered for segmented caches. A working-set approximation is developed for SIEVE as well as its length-aware extension.