Linear Elastic Caching: A New Approach to Cache Management Optimizes Cloud Costs
Google/DeepMind
Google Research and Google Cloud have introduced linear elastic caching, a method that dynamically adjusts cache size to minimize total cost. By framing page eviction as a ski rental problem and using a lightweight machine learning model, the approach reduced memory use by up to 30% in Spanner production servers with only a 0.5% increase in I/O costs.
Google Research and Google Cloud presented linear elastic caching at CIDR 2025. The method treats memory as a variable cost and uses a ski rental algorithm to determine time-to-live (TTL) for cached pages, combined with lightweight machine learning (a shallow decision tree) to predict optimal TTL. In Spanner production, it reduced memory consumption by up to 30% while causing a negligible 0.5% increase in I/O costs. Public trace evaluations also showed consistent cost savings over fixed-size caches. The work was done by Todd Lipcon, Manish Purohit, Tamas Sarlos, and Ravi Kumar.
- Сокращения
- CIDR = Conference on Innovative Data Systems Research — Конференция по инновационным исследованиям систем данных
- TTL = time-to-live — время жизни
- LRU = least recently used — наименее недавно использованный
- I/O = input/output — ввод/вывод
- ML = machine learning — машинное обучение
- TCO = total cost of ownership — совокупная стоимость владения
- GDSF = greedy dual size frequency — жадный двойной по размеру и частоте
Source: Google Research —
original