Caching for dataset-based workloads with heterogeneous file sizes

Abstract

Caching can effectively reduce the cost of serving content and improve the user experience. In this paper, we explore the benefits of caching for existing scientific workloads, taking the Worldwide LHC (Large Hadron Collider) Computing Grid as an example. It is a globally distributed system that stores and processes multiple hundred petabytes of data and serves the needs of thousands of scientists around the globe. Scientific computation differs from other applications like video streaming as file sizes vary from a few bytes to terabytes and logical links between the files affect user access patterns. These factors profoundly influence caches' performance and, therefore, should be carefully analyzed to select which caching policy to deploy or to design new ones. In this work, we study how the hierarchical organization of the LHC physics data into files and groups of files called datasets affects the request patterns. We then propose new caching policies that exploit dataset-specific knowledge and compare them with file-based ones. Moreover, we show that limited connectivity between the computing and storage sites leads to the delayed hits phenomenon and estimate the consequent reduction in the potential benefits of caching.

Data Popularity for Cache Eviction Algorithms using Random Forests

“…Another key distinction lies in the optimization goal. While most of the scientific research papers deal with cached objects of the same size and, therefore, focus on optimizing File Miss Ratio (FMR) [7], our data involves considerably large file sizes with a wide distribution (figure 1). Hence, we are primarily interested in optimizing Byte Miss Ratio (BMR) -a metric that takes into account the impact of file size on cache performance.…”

“…However, it raises the question of whether a better BMR can be achieved through the adoption of alternative cache eviction policies. To explore the performance boundaries, we examine PFOO-U.Bytes [7], representing a tighter lower bound for BHR than an infinite-size cache [9]. In figure 3 it is marked as "Optimum (Lower bound)".…”

“…In an attempt to close this gap, we have previously explored existing state-of-the-art policies, such as 2-LRU [8], MRU [10], GDSF [11], and a modification of AdaptSize [12] suitable for our case. Moreover, we have tailored cache eviction policies specific to our types of workloads, such as Dataset Evict LRU and Dataset Evict MRU [7], which leverage the observation that files within a dataset are often read together during analysis jobs. Despite our extensive exploration and efforts, we were unable to surpass the performance of LRU in previous attempts (see "2-LRU" and "Dataset Evict LRU" on figure 3 as an example).…”