Hacker News new | past | comments | ask | show | jobs | submit login
Rethinking SIMD Vectorization for In-Memory Databases [pdf] (columbia.edu)
69 points by luu on Jan 18, 2016 | hide | past | favorite | 1 comment



The paper did not mention it but it should also be noted that the bit-level parallel scatter and gather instructions (PDEP/PEXT) introduced with Haswell as part of the BMI2 extensions are also quite useful for high-performance database implementations and can be used similarly to the word vector scatter/gather instructions.

Everything mentioned in the paper will work just as well for SSD-backed databases as pure in-memory databases assuming competent design and implementation. The in-memory constraint allowed an "apples to apples" comparison using the Xeon Phi coprocessor, which does not have direct access to storage.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: