Add GPU RFC

RFC-0021-gpu-support-cudf.md

@@ -0,0 +1,68 @@
+# **RFC-0021 GPU support using cuDF in C++ workers**
+
+### Proposers
+
+* Deepak Majeti, et. al. (IBM)
+* Zoltan Arnold Nagy, et. al. (IBM Research Europe)
+* Karthikeyan Natarajan, et. al. (NVIDIA)
+
+## Related Issues
+
+* [PR #25094](https://github.com/prestodb/presto/pull/25094): Enable Velox cuDF
+* [PR #26156](https://github.com/prestodb/presto/pull/26156): Add support for Velox cuDF options and CudfHiveConnector
+
+## Summary
+
+Enable C++ workers to execute queries on GPUs. 
+
+## Background
+
+There is now a proliferation of GPU hardware primarily due to the demands from AI/ML usecases.
+GPU hardware over the years has evolved with advanced I/O capabilities.
+New AI adjacent data processing workflows are also being developed.
+
+GPUs provide high compute and memory bandwidth, which can benefit operations such as
+joins, aggregations, string processing, etc.
+
+
+### Goals
+* Allow Presto queries to run on a single GPU or multiple GPUs.
+* A query will run either on the CPU or a GPU. No hybrid execution.
+* Use CPU if a GPU lacks a certain functionality.
+* Execution should maximize utilization of available hardware such as NVLink.
+
+## Proposed Implementation
+
+Some of this work has been implemented in [Velox](https://github.com/facebookincubator/velox/tree/main/velox/experimental/cudf).
+The current implementation translates the CPU operators to the GPU operators via a DriverAdapter in Velox.
+
+Nvidia's [blog](https://developer.nvidia.com/blog/accelerating-large-scale-data-analytics-with-gpu-native-velox-and-nvidia-cudf/)
+has more details on the design and some early results.
+
+The [Extending Velox - GPU Acceleration with cuDF](https://velox-lib.io/blog/extending-velox-with-cudf) blog also covers the current implementation.
+
+On the Presto C++ side, the following registrations and configs have been added.
+
+* CMake build option `PRESTO_ENABLE_CUDF` must be set. https://github.com/prestodb/presto/tree/master/presto-native-execution#nvidia-cudf-gpu-support
+* Parquet file-format is supported. cudfHiveConnector is registered.
+* S3 and local/linux filesystems are supported.
+* cuDF [configs](https://facebookincubator.github.io/velox/configs.html#cudf-specific-configuration-experimental) can be 
+specified inside `config.properties` and catalog `.properties` file.
+
+The current work so far shows that GPUs can provide good price-performance. However, to make this support user-friendly and get better price-performance, the following improvements are in progress.
+
+## Work in Progress
+* Add GPU plan nodes.
+    * Driver adapter runs after the drivers/pipelines are built. Limits the adaptation.
+    * Allow efficient fallback to CPU.
+* GPU-GPU exchange using UCX (https://github.com/prestodb/presto/tree/ibm-research-preview).
+* Topology and hardware detection.
+* Metadata queries on CPU only.
+* Session parameter to filter workers.
+* Optimizer cost model to support GPUs.
+
+## Releases
+Presto C++ workers will be released with GPU support.
+
+## Test Plan
+Velox CI has a gpu runner sponsored by Meta. We need a similar runner for Presto.