Turn of the Carousel - What Does Edge Computing Change for Distributed Applications?

Online Learning for Network Resource Allocation

“…But object popularity can be difficult to estimate and can change over time, specially at the level of small geographical areas (as in the case of areas served by an edge server) [166]. Other papers [118][119][120][121][122]167] present more a high-level view of the different components of the application system, without specific contributions in terms of cache management policies (e.g., they apply minor changes to exact caching policies like LRU or LFU). Some recent papers [127,168,169] propose online caching policies that try to minimize the total cost of the system (the sum of the dissimilarity cost and the fetching cost), also in a networked context [168,170], but their schemes apply only to the case 𝑘 = 1, which is of limited practical interest.…”

“…Likewise, recommender systems can leverage similarity caches [110,128]: a recommender system can save operating costs and decrease its response time through recommendation of relevant contents from a cache to user-generated queries, i.e., in case of a cache miss, an application proxy (e.g., YouTube) running close to the helper node (e.g., at a multiaccess edge computing server) can recommend the most related files that are locally cached. More recently, similarity caches have been employed extensively for machine learning based inference systems to store queries and the respective inference results to serve future requests, for example, prediction serving systems [117], image recognition systems [118,119,121], object classification on the cloud [122], caching hidden layer outputs of a neural network to accelerate computation [123], network traffic classification tasks [145]. The cache can indeed respond with the results of a previous query that is very similar to the current one, thus reducing the computational burden and latency of running complex machine learning inference models.…”

“…Similarity search [104] is a key building block for a large variety of applications including multimedia retrieval [105][106][107], recommendation systems [108][109][110], genome study [111,112], machine learning training [113][114][115], and serving [116][117][118][119][120][121][122][123][124][125]. Given a query for an object, the goal is to retrieve one or more similar objects from a repository.…”

“…This operation was named similarity caching, in contrast to the traditional exact caching. As recognized in [127], the idea of similarity caching has been rediscovered a number of times under different names: recognition caches [118,119], approximate deduplication [121], semantic caches [122], prediction caches [117], approximate caches [123], and soft caches [110,128].…”

“…Examples of commonly employed distances are the 𝑝-norm, Mahalanobis or cosine similarity distances. For instance, in augmented reality applications we often require identifying similar objects: the image is coded into a query, i.e., an embedding in R 𝑑 (e.g., set of descriptors like SIFT [129], or ORB [130], or the set of activation values at an intermediate layer of a neural network [131,132]), and the application logic finds similar images to be returned to the user [118][119][120]122]; moreover, in recommendation systems, objects are also embedded in R 𝑑 with a distance metric to capture content dissimilarity. This also the case for other potential applications of similarity caching like 360 • videos, where tiles at the periphery of the user's field of view could be approximated by neighboring tiles that are stored at the cache and can then be served with low latency [133].…”