Paper Title: Why cache in DRAM if you use NVMe-oF

Paper Abstract: Many applications require fast access to data stored in highly available but slow storage systems. This requirement is usually achieved by caching data in DRAM in front of the storage system using a system like Memcached or Redis. Such an approach is, however, not ideal. DRAM is both limited in size and expensive which results in low per server data density and high cost per gigabyte. We propose replacing DRAM caches with fast NVMe-oF connected storage. We show that a configuration composed of several caching servers running uDepot on top of an All Flash Array delivers the DRAM-equivalent performance, while offering much higher caching capacities at a dramatically reduced cost. uDepot is a key-value store built to avoid all inefficiencies of traditional I/O stacks and to enable applications to use fast storage at its full potential. This configuration also disaggregates storage from compute and allows dynamically scaling the number of caching servers to match application needs.

Paper Author: Radu Stoica, Researcher, IBM

Author Bio: Radu Stoica is a Research Staff Member part of the Cloud and Computing Infrastructure at IBM Research Zurich. He is broadly interested in storage related topics, with an emphasis on storage class memories and cloud applications. His current work includes designing enterprise-grade flash controllers, developing data deduplication techniques, improving the efficiency of strongly consistent distributed databases and leveraging in-memory distributed filesystems as accelerators. Before joining IBM, he completed his Ph.D. at EPFL working in the DIAS group under the supervision of Prof. Anastasia Ailamaki. His thesis focused on adapting traditional relational database systems to take full advantage of new storage hardware such as flash memory.