Remote object stores are widely used in cloud workloads for their high availability, redundancy, and elastic scalability, which allow users to offload infrastructure management. However, accessing these remote stores introduces additional latency and cost. To mitigate these drawbacks, local caching layers are often employed to improve performance and reduce request overhead. While DRAM-based caches offer high performance, they are expensive and impractical for large datasets. Flash-based (SSD) caches provide a more cost-effective alternative, but their performance can degrade under heavy workloads due to background garbage collection (GC) processes inherent to conventional SSD architectures. Zoned Namespaces (ZNS) SSDs have recently emerged as a promising alternative to traditional flash storage. By eliminating the need for device-level GC, ZNS SSDs can provide more predictable performance. However, they impose new constraints on data placement and require significant changes in cache design and management. In this work, we design and implement a ZNS-aware caching system to evaluate the performance trade-offs between ZNS and traditional block-interface SSDs in caching workloads. Through experimental analysis, we identify when ZNS devices can outperform conventional SSDs, and when the benefits may not justify the required software complexity. We conclude by providing design recommendations for caching system developers considering the adoption of ZNS SSDs.
John Ramsden (Thu,) studied this question.