Generalized Che's Cache Approximation

• Generalization of Che's approximation is an analytical framework that extends the decoupling principle for evaluating diverse caching policies and network configurations.
• It incorporates factors such as temporal locality, renewal traffic models, and multi-stage cache networks while preserving computational efficiency.
• Validation through LLN, CLT, and LDP guarantees high accuracy and scalability in predicting cache hit rates in complex, modern systems.

The generalization of Che's approximation refers to a broad class of analytical techniques that extend the original Che approximation—introduced as a decoupling scheme for isolating interactions in performance analysis of systems such as caches—across a variety of policies, traffic models, and network topologies. This generalized framework underpins modern high-fidelity yet computationally lightweight cache analysis and has become foundational for both theoretical and applied work in caching systems.

1. The Original Che’s Approximation and Decoupling Principle

Che's approximation originated as a method for analyzing isolated Least Recently Used (LRU) caches under the Independent Reference Model (IRM) where each object receives requests as an independent Poisson process. The central insight is the decoupling of eviction dynamics: the random cache eviction time for each object, $T_C(m)$, is approximated by a single deterministic characteristic eviction time $T_C$ for the entire cache. This reduces the analysis to solving a fixed-point equation:

$$C = \sum_{m=1}^{M} \left[1 - e^{-\lambda_m T_C}\right]$$

where $C$ is cache capacity, $M$ is the catalogue size, and $\lambda_m$ is the request rate for object $m$. The hit probability for each object is then $1 - e^{-\lambda_m T_C}$, and the global hit rate aggregates over all objects. This leads to an $O(M)$ per-iteration complexity and delivers highly accurate predictions for a wide range of catalogue sizes and popularity distributions (Martina et al., 2013).

2. Extension to General Caching Policies

The generalization covers a rich family of caching policies, including but not limited to q-LRU (probabilistic insertion), FIFO, RANDOM, and $k$-LRU hierarchies:

• q-LRU: The probabilistic insertion mechanism leads to nonlinear balance equations for hit/in probabilities. The approximation yields:

$T_C$0

This extends to the renewal context by replacing the exponential terms with general inter-arrival CDFs (Martina et al., 2013).

• FIFO/RANDOM: For FIFO and RANDOM, the cache residency probability is governed by the M/G/1/0 (loss-system) formula:

$T_C$1

where $T_C$2 is determined self-consistently (Martina et al., 2013).

• $T_C$3-LRU: In multi-stage cache hierarchies, the approximation is recursively extended, with each stage’s eviction time parameterized and solved via coupled fixed-point equations, converging to LFU as $T_C$4 (Martina et al., 2013).

3. Incorporating Temporal Locality and Renewal Traffic

In practical systems, traffic deviates from the Poisson assumption due to temporal locality. The generalization accommodates renewal traffic: every object’s request process is replaced by a renewal process with an arbitrary inter-request distribution $T_C$5. The residence and hit probabilities are computed through the age and inter-arrival CDFs:

$T_C$6

The size constraint and all fixed-point equations are reformulated in terms of these general distributions, preserving the computational efficiency and advancing prediction accuracy, particularly for workloads exhibiting burstiness or temporal clustering (Martina et al., 2013).

4. Extension to Cache Networks

The decoupling principle underlying Che’s approximation is further generalized to interconnected cache networks (tandem, tree, mesh) by tracking miss streams and conditional hit probabilities along the network graph:

• Improved Miss-Stream Modeling: Rather than using a naive Poisson assumption for miss streams at downstream caches (which severely misestimates network-level hit rates), the generalized Che analysis computes the actual sojourn window resulting from upstream cache eviction times.
• Window-based Decoupling: For a tandem of two caches, the probability that an object is resident in cache 2 is expressed as

$T_C$7

if $T_C$8, where $T_C$9 is the rate of misses passed from cache 1. This corrects the systematic overestimation of naive models (Martina et al., 2013).

• General Networks: For arbitrary network topologies, the relevant window and miss-fraction parameters are propagated recursively, with each edge’s behavioral model precisely capturing detailed traffic dependencies (Martina et al., 2013).

5. Asymptotic and Probabilistic Validation

Rigorous large-system validation accompanies the generalization:

• Law of Large Numbers (LLN): Under suitable continuity assumptions, the actual eviction time concentrates around the deterministic $$C = \sum_{m=1}^{M} \left[1 - e^{-\lambda_m T_C}\right]$$0 in the large-cache limit.
• Central Limit Theorem (CLT): The fluctuation of the eviction time about its deterministic approximation admits a Gaussian scaling with explicit variance parameters, subject to the regularity of the key function $$C = \sum_{m=1}^{M} \left[1 - e^{-\lambda_m T_C}\right]$$1 (Leonardi et al., 2014).
• Large Deviations Principle (LDP): The probabilities of large deviations in eviction time obey exponential tail estimates with rate function $$C = \sum_{m=1}^{M} \left[1 - e^{-\lambda_m T_C}\right]$$2, yielding exponential bounds for rare eviction events (Leonardi et al., 2014).
• Analytical Bounds in Networks: For tandem networks, explicit upper and lower bounds can be derived via inclusion-exclusion and combinatorial analysis of the windowed intervals, improving precision in the approximation (Leonardi et al., 2014).

6. Computational and Practical Implications

The generalized Che approximation preserves linear or near-linear computational complexity in all scenarios:

• Efficiency: Each fixed-point iteration remains $$C = \sum_{m=1}^{M} \left[1 - e^{-\lambda_m T_C}\right]$$3 for object count and $$C = \sum_{m=1}^{M} \left[1 - e^{-\lambda_m T_C}\right]$$4 for network size ($$C = \sum_{m=1}^{M} \left[1 - e^{-\lambda_m T_C}\right]$$5 number of edges) (Martina et al., 2013).
• Accuracy: Empirical validation demonstrates accuracy within a few percent of full-trace simulation even for large networks (e.g., trees with thousands of caches) (Martina et al., 2013).
• Scalability: The extension to generalized network, policy, and traffic scenarios enables tractable analysis for modern content-delivery infrastructures.

7. Design Insights and Trade-offs

The unified framework uncovers key practical conclusions:

• Policy Superiority: Insertion-filtering policies ($$C = \sum_{m=1}^{M} \left[1 - e^{-\lambda_m T_C}\right]$$6-LRU, $$C = \sum_{m=1}^{M} \left[1 - e^{-\lambda_m T_C}\right]$$7-LRU, LCD) outperform FIFO/RANDOM, especially for small caches.
• Locality Sensitivity: Temporal locality (modeled via renewal traffic) can more than double hit rates vs. Poisson models, emphasizing the need for locality-aware analysis.
• Network Effects: Accurate modeling of cache interactions is crucial for replicative and hierarchical cache networks—naive approximations lead to substantial mispredictions (Martina et al., 2013).

These insights drive both the theoretical and applied development of caching architectures and traffic-shaping strategies, demonstrating the critical value of generalized Che-style approximations for system design and analysis.

References:

(Martina et al., 2013, Leonardi et al., 2014)