Hash tables are an essential data-structure for numerous networking applications (e.g., connection tracking, firewalls, network address translators). Among these, cuckoo hash tables provide excellent performance by allowing lookups to be processed with very few memory accesses (2 to 3 per lookup). Yet, for large tables, cuckoo hash tables remain memory bound and each memory access impacts performance. In this paper, we propose algorithmic improvements to cuckoo hash tables allowing to eliminate some unnecessary memory accesses; these changes are conducted without altering the properties of the original cuckoo hash table so that all existing theoretical analysis remain applicable. On a single core, our hash table achieves 37M lookups per second for positive lookups (i.e., when the key looked up is present in the table), and 60M lookups per second for negative lookups, a 50% improvement over the implementation included into the DPDK. On a 18-core, with mostly positive lookups, our implementation achieves 496M lookups per second, a 45% improvement over DPDK.

13 pages