Build:Divide R into chunksConstruct a bloom filter for each chunkProbe Join:For each value in S:

Probe all Bloom filters and perform join for qualifying chunks

Evaluation

Radix Hash

4567

0x 0000000123

RH

0x 0000000123

Bloom filter joins in NUMA Hash joins in NUMA

BFJ in modern hardwareBloom filter internals

Using radix hashes improved performance without impacting accuracy As opposed to hash joins, Bloom filter joins scale linearly in

NUMA environments

Joins with low selectivity

Bloom Filters are flexible and light-weight structures

Hash tables require multiple passes, concurrency control and more space

30%

70%

Join Other

Motivation

Bloom Filter JoinsAbdul Wasay, Manos Athanassoulis, Stratos Idreos

Comparing Bloom Filters

Bloom Filter Join

Preliminary ResultsNUMA: Non-uniform Memory access

BFJ perform comparably to hash joins for smaller size of S

Performance comparison of various implementations of BFJs vs. number of BF

1

2

For higher selectivity the performance of BFJ degrade very fast