Hash Table for Large Key-Value Stores on FPGAs
Zsolt István · Repository for Publications and Research Data (ETH Zurich) · 2013
Hash tables gain increasing importance in big data applications such as memcached.In software hash tables can be easily sized to hold millions of entries, and collisions are most often solved by chaining techniques.As opposed to software hash tables, hardware implementations are more challenging as the unpredictable lookup latencies complicate system-level architectures.In this work we present a novel hash table architecture that is fully pipelined to offer high performance for main memory key-value stores used as caches.Key contributions include support for flexible key sizes, collision resolution through parallel lookup and a scalable architecture that can accommodate for well more than 1 million entries stored in off-chip DRAM.We also discuss potential cache management strategies which exploit the advantages of hybrid CPU-FPGA platforms.