Approximate Computation Caching in Information Centric Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Information Centric Network (ICN) and Named Function Networking (NFN) mechanisms do not effectively reuse approximate computation results when input data varies, leading to additional computational overhead and unnecessary network traffic due to exact matching requirements.
Innovation Solution
Implementing an approximate computation cache in ICN devices that uses similarity-based matching for NFN requests, employing locality sensitive hashing and ontology tagging to store and retrieve computation results based on feature sets and similarity scores, allowing for the reuse of computation results even when input data names differ.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If exact matching is used between interest packet names and cached computation results, then caching reliability is improved, but cache hit rate deteriorates when input data varies
Solution Approach 1:
The patent changes the matching parameter from exact name matching to similarity-based matching using feature sets. Instead of requiring identical interest packet names, the system extracts features from input data and performs approximate matching, allowing cached results to be reused even when input data varies slightly. This resolves the contradiction by maintaining reliability through structured feature comparison while improving cache hit rate through flexible similarity matching.
Solution Approach 2:
The patent applies different matching strategies to different parts of the caching system. Exact matching is maintained for the interest packet name structure, while approximate matching based on feature set similarity is applied to the computation result retrieval. This dual approach ensures both reliability in cache organization and improved hit rate in result retrieval.
2Productivity
If computation results are cached and reused, then computational load is reduced, but network traffic increases due to additional caching operations
Solution Approach 1:
The patent performs preliminary extraction of feature sets from input data and stores them alongside computation results in the cache. When a new computation request arrives, the feature set is extracted and compared against cached feature sets before triggering actual computation. This preliminary action enables early identification of reusable results, reducing the need for actual computation and associated network traffic.
Solution Approach 2:
The patent creates feature set representations (copies) of input data that can be stored and compared without requiring the original large-scale data. These feature set copies enable efficient cache lookup and comparison operations, reducing the computational overhead and network traffic associated with moving and processing full input datasets.
3Adaptability or versatility
If approximate matching is implemented for computation results, then cache versatility is improved, but matching precision deteriorates
Solution Approach 1:
The patent implements a dynamic matching threshold mechanism where the similarity threshold can be adjusted based on the specific computation task and input data characteristics. For computations requiring high precision, the threshold is set stricter, while for more tolerant applications, a more relaxed threshold increases cache versatility. This dynamic adjustment resolves the contradiction by adapting the matching precision to the specific use case.
Solution Approach 2:
The patent replaces traditional exact string-matching mechanisms with feature-based similarity comparison using cosine similarity or other distance metrics. This substitution enables approximate matching that captures semantic similarity between different input representations, significantly improving cache versatility while maintaining acceptable precision through mathematical similarity measures rather than rigid exact matching.
Data Source
AI summary
Systems and techniques for information centric network (ICN) approximate computation caching are described herein. For example, an interest packet that includes a feature set of input data may be received. A node may then perform a search of a local data store using the feature set to determine an approximate computation result cached in the local data store. Here, the approximate computation result may be based on input data that differs from the input data named in the interest packet. The node may then return the approximate computation result to an author of the interest packet in response to the search.