HTM Hash Join

Implements an adaptive multi-core hash join based on hardware transactional memory (HTM-Adaptive).

The join does > 2x better than a radix join (PRJ) on data with locality and with low-overhead switches to radix join in the absence of locality.

For details, see paper link

Also contains implementations of:

• hash join using a global c++ 11 atomics-based hash table
• parallel radix join based hash join (PRJ)
• hash join using a global hash table which uses locks (NPJ)

Installation

Note that running HTM hash join requires hardware that supports Hardware Transactional Memory via Intel TSX. Use this tool to check if your hardware has TSX.

Make sure you have Intel TBB downloaded.

cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_MODULE_PATH=/path/to/FindTBB
make 

A binary main created is used to run all the different algorithms.

./main --algo <algo>
       --rSize <size of build side>
       --transactionSize <number of hash table inserts per transaction>
       --dataDistr <data distribution>
       --shuffleRange <shuffle window (only used if dataDistr == local_shuffle)>
       --numPartitions <number of partitions>

algo = htm, nocc, atomic
dataDistr = uniform, sorted, shuffle, local_shuffle

Contact

In case of issues, file an issue or email: anils [at] mit [dot] edu