I build systems.
1. I do care about GPU systems and their future, but that does not mean I care any less about CPU systems.
2. I do open research. I make everything open whenever possible.
3. I do care about industry applications and highly value research that can be useful in industry to ultimately benefit people. I build systems for people to use, not only for academic records or novelty.
4. I do my best to write educational papers. In this effort, I feel responsible for openly sharing my insights with others building similar systems.
My Research Impact
Inspired and encouraged by several people, including Xiangpeng Hao and Peter Boncz, I define research impact as building practical systems that solve real-world problems, benefit people, and advance the public good.
Here, I collect examples of the impact of my work:
- PystachIO
- Parquet reader in libcudf: I identified a performance issue in the GPU Parquet reader and developed a complete POC to address it. I merged the first half of the POC into libcudf, and the libcudf team carried the remaining work forward. (I am deeply grateful to Muhammad Haseeb and Vukasin Milovanovic for their help!): my libcudf story issue.
- Broader improvements beyond Parquet: The issue I identified was not specific to Parquet, but part of a broader challenge in data-intensive applications. The cuDF team and I called for a rewrite, leading to improvements beyond Parquet: the know-how, the fix in cuDF, the fix in cuCollections.
-
As libcudf is an underlying library, all systems that depend on it can benefit from my work, including cuDF-Spark, Velox-cuDF, Presto-cuDF, Sirius, GQE, and others.
- Do GPUs Really Need New Tabular File Formats?
-
This work is a call to action not only for the research community, but also for the Parquet and Iceberg communities: my paper incorporated feedback from Andrew Lamb (Parquet PMC), mentioned by Gang Wu (Parquet PMC) in his talk.