Efficient Hash Probes on Modern Processors

Bucketized versions of Cuckoo hashing can achieve 95-99% occupancy, without any space overhead for pointers or other structures. However, such methods typically need to consult multiple hash buckets per probe, and have therefore been seen as having worse probe performance than conventional techniques for large tables. We consider workloads typical of database and stream processing, in which keys and payloads are small, and in which a large number of probes are processed in bulk. We show how to improve probe performance by (a) eliminating branch instructions from the probe code, enabling better scheduling and latency-hiding by modern processors, and (b) using SIMD instructions to process multiple keys/payloads in parallel. We show that on modern architectures, probes to a bucketized Cuckoo hash table can be processed much faster than conventional hash table probes, for both small and large memory-resident tables. On a Pentium 4, a probe is two to four times faster, while on the Cell SPE processor a probe is ten times faster.