Independent reference model
The independent reference model describes cache requests as independent draws from a fixed popularity distribution over a fixed catalogue.
The request sequence has no temporal correlations. The next identity does not depend on earlier identities. Independent homogeneous Poisson streams provide a continuous-time representation with request probabilities proportional to their rates. [1]
Measured marginal popularities alone do not establish independence. For example, repeating ABAB and AABB both request A and B equally often. With one LRU slot their long-run hit ratios are respectively 0 and 1/2, while independent equal-probability draws give 1/2. The shared frequency histogram is insufficient to predict cache behaviour. These are constructed controls, not workload measurements. [2]
The Che approximation uses this assumption in its familiar exponential form. A model that matches popularity can still miss ordering, bursts or changes in the catalogue. Preserve the observation layer and distinguish steady-state estimates from a finite trace with an empty initial cache. [1]
Relations
- related
- Che approximation, Least recently used replacement, Miss-ratio curve, Locality of reference, Tsetlin library
Citations
- Martina, Garetto and Leonardi. A unified approach to the performance analysis of caching systems. arXiv:1307.6702v6 (2016), section II.A.
- Hesela. Exact periodic one-slot controls and independent-reference arithmetic.
Sources & caveats
- A workload assumption, not a replacement algorithm. Periodic controls are analytically constructed and not samples drawn from IRM.