KRCore has two parts. The first is a rust-based library for using RDMA, i.e., by adding high-level abstractions to low-level RDMA libraries (ibverbs). The second is a kernel-space component to provide fast user-space RDMA-capable connection setup, which builds on the first part.
[TOC]
Crate location: ./KRdmaKit
.
The goals of this part is:
- Simplify building RDMA-based applications with rust.
- Using rust to manage all the (local) lifecycles of RDMA resources (e.g., QP) in a safe and transparent way.
- A unified codebase to build RDMA-enabled applications both in the kernel-space and user-space.
Crate location: ./KRdmaKit-syscall
.
The second is a kernel-space component to provide fast user-space RDMA-capable connection setup. For details, please refer to our USENIX ATC 2022 paper: https://www.usenix.org/conference/atc22/presentation/wei .
To use this part or re-produce evaluations in our paper, please go to https://github.com/SJTU-IPADS/krcore-artifacts/tree/atc22-artifacts, or git checkout atc22-artifacts
.
Please refer to install.md for the environment setup and exp.md for the evaluations. All documents can be found in the doc
directory of the artifact.
Since KRCore is a kernel-space solution with RDMA, we should first build KRCore from source at the machines involved in the evaluations. Afterward, we provide scripts so that all the evaluations can be done on a single controller machine.
- support user-space RDMA programs with the same KRdmaKit crate
- build more utilities (e.g., reliable messaging over UD) atop of KRCore
- revise system call layer to the new KRdmaKit API
- support cross-communication between user and kernel
Overview of roles of machines involved in the evaluations
At high-level, we need one machine to be KRCore server (that serves RDMA control path and data path operations), a varied number of machines to be KRCore clients (that issues RDMA control path and data path operations to the server), and one controller machine that conduct experiments by issuing commands to the clients and server. The server and clients should connect with RDMA. The controller can be any machine that can establish ssh connections to the clients and server.
Except for building KRCore dependncies at clients and servers, all AE can be finished via scripts at the controller.
To get started, please first select a controller machine (e.g., your laptop or a PC in the lab).
At the controller. clone the repo to some directory (/path/to/krcore
) and export it:
git clone https://github.com/SJTU-IPADS/krcore-artifacts.git
export PROJECT_PATH=/path/to/krcore
At the controller. Then fetch the third-party dependencies.
cd ${PROJECT_PATH}
sh git_init.sh
At the controller. Then synchronize all the files to all the hosts* of clients and server. We have prepread a sync script (rsync.sh
) shown below that you can use for the synchronization.
user="your_user_name"
target=("host1" "host2") ## all the client and server hosts
path="./krdmakit"
for machine in ${target[*]}
do
rsync -i -rtuv \
$PWD/deps \
$PWD/KRdmaKit-syscall \
$PWD/rust-kernel-rdma \
$PWD/krdmakit-macros \
$PWD/testlib \
$PWD/include \
$PWD/mlnx-ofed-4.9-driver \
$PWD/exp \
$user@$machine:/home/${user}/${path} \
done
Important!!! Please ensure each host of the clients and server has the same login username and password (i.e. user="test_user", pwd="my_pwd" for each host). Also, the path of the KRCore base is the same at all of them. For example, in the above rsync.sh
, all the KRCore code are synchronized to /home/${USER}/projects/krdmakit
at the clients and server.
Afterwards, please go to the next section for installing KRCore on a specific host.
See install.md (in ${PROJECT_PATH}/doc
)for installing KRCore on a single machine. All the hosts involved should follow the same steps.
After finish the install instructions in install.md on all the involved servers, we can now start to reproduce results of KRCore shown below.
If Check if the configuration is correct in Run Experiments of exp.md passes, then everything will be fine. Please refer to exp.md for more details.
There are three key results of the evaluations of the paper:
-
KRCore can significantly improve the performance of RDMA control path (compared with verbs), e.g., 1000X+ higher throughput.
Corresponding paper figures: Figure 8a
-
KRCore have a comparable RDMA data path performance (compared with verbs) for common RDMA operations, e.g., one-sided RDMA and two-sided RDMA.
Corresponding paper figures: Figure 10-11
-
KRCore can improve the application (RaceHashing) bootstrap time when starting and connecting new nodes for the computation.
Corresponding paper figures: Figure 14
We focus on the reproducing of these figures---others are similar in principle (or is less relevant to the key results), but is challenging for automatic reproduction due to configuration issues. For how to reproduce the key result figures, please refer to the next section.
See exp.md for more information. (in ${PROJECT_PATH}
)