Our project explores the implications of replacing MICA with 2 tree based key-value stores. We have used a B+-tree implementation, and SplinterDB.
This document specifies instructions on how to set the system up on an RDMA cluster and run experiments.
| Name | Contributions |
|---|---|
| Aaditya Ranga Rajan | SplinterDB Integration |
| Ankith Boggaram | B+-tree integration |
| Soham Bagchi | Cluster setup, metrics evaluation and analysis |
Our experiments were run on cloudlab. We used 5 nodes of the r320 class, each with a Mellanox NIC with RDMA capabilities. .
In order to setup the project, please run the init-preimaged.sh script available in the repo. It will download the necessary dependencies such as CMake, and will also create an ext4 partition on your device. Please make sure to edit the paths mentioned in the script.
Once this is done, please change the IPs in bin/cluster.sh. These should be the IPs of your nodes. Also, there is a variable called $OD_HOME. Please set this to your Odyssey directory.
On your main node, please run the following commands
cmake -B build
./bin/copy-run.sh -x hermes
This should build the executable and copy it to all the other nodes in your cluster. Now, on your other nodes, execute this
./bin/run-exe.sh -x build/hermes
Once you have executed this command on all the nodes successfully, Odyssey should be up and running.It will proceed to run YCSB workload A. This runs MICA as the KVS by default. If you would like to run B+-tree or SplinterDB, please make the following change.
In odlib/include/general_util/od_top.h, (lines 92 - 94), you will find 3 macros.
#define USE_MICA 1
#define USE_BPLUS 0
#define USE_SPLINTERDB 0
If you want to use SplinterDB, set that macro to 1. Please remember that only one of these can be set to 1 at a time, or else the code will not run.
There is still room for improvement in the code. The existing Odyssey code is not the best in terms of documentation, and that can certainly be improved. The paths in the code currenly are hardcoded. We plan to make it such that we can get all required paths from environment variables. A direction for potential research is integration an adaptive Be-tree and running the experiments again. Feel free to contribute!