SMACC

As cloud computing is gaining popularity among small and medium enterprises, cloud storage solutions such as Amazon S3 are increasingly utilized for storing, maintaining, and serving application data. Despite the typically high-speed internet connections between applications and Cloud storage, there is still a huge performance gap compared to accessing data from direct-attached memory or even locally attached disks. SMACC is a novel multi-tier Cloud caching service developed at CUT that can run on application compute nodes (e.g., on Amazon EC2 or on premise) and cache frequently-used data residing on cloud storage (e.g., Amazon S3, MinIO) into local memory and locally-attached disks (e.g., Amazon EBS) using new smart policies. SMACC also provides an HDFS-compatible API interface, which can be used by big data platforms such as Spark and Hadoop for processing data residing on Amazon S3, while caching data blocks on the various compute nodes for increased performance.

RLCM, a reinforcement learning (RL)-based cache management framework has been integrated with SMACC for making all cache-related decisions, including initial data placement, data upgrades to higher tiers, data downgrades to lower tiers, and data evictions from tiers. RLCM contains (1) an Admission Agent tasked with selectively admitting and upgrading the most relevant and high-priority data into the cache tiers, ensuring an optimal balance between cache utilization and performance; and (2) an Eviction Agent tasked with strategically evicting less relevant or infrequently accessed items while retaining valuable ones when a cache tier reaches capacity, as well as deciding whether to downgrade or remove the evicted items from the cache. Finally, the RL Manager component oversees the overall process of collecting request information from the caching system, storing it in a Metadata Repository for computing the environment states, managing the Admission and Eviction Agents, collecting usage statistics from the cache tiers, as well as computing rewards to update the reinforcement learning models.