t1mm3 / weld_tpch

Weld vs. vectorization & data-centric compilation

Comparison of Weld vs. state-of-the-art query paradigms

This repository is based on https://github.com/TimoKersten/db-engine-paradigms.

Building

Download sandbox.sh and modify BASE (base prefix to use) as well as N (number of threads).

Notes

Q1

• Already did this one (partially) as part of [1]
• A generalized group-by can be implemented using a dictmerger [2]

Q3

• Joins can be implemented using
• As Weld does not support strings, we need an UDF for string equality
• TimoKersten's string implementation requires us to store the length too i.e. a string is a tuple {i16, i64}

Q6

• Part of Weld benchmarks [3]

Q9

• GroupBy on string key, not clear how to do that, tried but will segfault
• We use a string dictionary now

References

• [1] Tim Gubner and Peter Boncz, Exploring Query Execution Strategies for JIT, Vectorization and SIMD, ADMS 2017, https://t1mm3.github.io/assets/papers/adms17.pdf
• [2] Mihai Varga. Just-in-time Compilation in MonetDB with Weld. Vrije Universiteit Amsterdam, 2018, https://homepages.cwi.nl/~boncz/msc/2018-MihaiVarga.pdf
• [3] https://github.com/weld-project/weld-benchmarks