Canvas is a remote-memory data path designed to provide resource isolation for cloud applications co-running on the memory disaggregated cluster and enable the adaptive management policies for each application according to its program semantics. This repository contains the source code of Canvas client ($repo_home_dir/remoteswap/client), and Canvas server ($repo_home_dir/remoteswap/server), the configuration for the Canvas scheduler ($repo_home_dir/remoteswap/configuration) and the Linux kernel that Canvas needs (linux5.5). Please refer to our NSDI'23 paper, Canvas, for more details.
Canvas is tested under the following setting(s):
Ubuntu 18.04/20.04
Linux 5.5
GCC 7.5.0/9.4.0
MLNX_OFED driver 5.0-2.1.8.0
Note that higher versions of MLNX-OFED driver (e.g., 5.8) may not support Linux 5.5.
Next we will use Ubuntu 20.04 as an example to show how to build and install the kernel. It is not required but highly recommended to have the same kernel version for both CPU and memory server.
(1) Change the grub parameters (at least on CPU server)
sudo vim /etc/default/grub
# Choose the bootup kernel version as 5.5.0-canvas
GRUB_DEFAULT="Advanced options for Ubuntu>Ubuntu, with Linux 5.5.0-canvas"
# Change the value of GRUB_CMDLINE_LINUX to set transparent hugepage as madvise:
GRUB_CMDLINE_LINUX="transparent_hugepage=madvise"
# Apply the change
sudo update-grub(2) Build the Kernel source code && install it.
# Change to the kernel folder:
cd Canvas/linux5.5
# In case new kernel options are prompted, press enter to use the default options.
cp config .config
sudo ./build_kernel.sh build
sudo ./build_kernel.sh install
# (optional) install built kernel only.
# sudo ./build_kernel.sh replace
sudo rebootPreparations:
Canvas is only tested on MLNX_OFED-5.0-2.1.8.0. Download and unzip the package according to your system version, on both CPU and memory server.
Take Ubuntu 20.04 as an example:
# Download the MLNX OFED driver for the Ubuntu 20.04
wget https://content.mellanox.com/ofed/MLNX_OFED-5.0-2.1.8.0/MLNX_OFED_LINUX-5.0-2.1.8.0-ubuntu20.04-x86_64.tgz
tar xzf MLNX_OFED_LINUX-5.0-2.1.8.0-ubuntu20.04-x86_64.tgz
cd MLNX_OFED_LINUX-5.0-2.1.8.0-ubuntu20.04-x86_64
# Remove the incompatible libraries
sudo apt remove ibverbs-providers:amd64 librdmacm1:amd64 librdmacm-dev:amd64 libibverbs-dev:amd64 libopensm5a libosmvendor4 libosmcomp3 -y
# Install the MLNX OFED driver against the kernel 5.5.0
sudo ./mlnxofedinstall --add-kernel-supportDisclaimer: This step is only required if you are using InfiniBand and want to configure your own subnet. Operations below are what we did in our small internal research-oriented cluster. Make sure you understand what you are doing before executing commands below.
(1) Enable and start the openibd service
sudo systemctl enable openibd
sudo systemctl start openibd
# confirm the service is running and enabled:
sudo systemctl status openibd
# the log shown as:
● openibd.service - openibd - configure Mellanox devices
Loaded: loaded (/lib/systemd/system/openibd.service; enabled; vendor preset: enabled)
Active: active (exited) since Mon 2022-05-02 14:40:53 CST; 1min 24s ago
(2) Enable and start the opensmd service:
sudo systemctl enable opensmd
sudo systemctl start opensmd
# confirm the service status
sudo systemctl status opensmd
# the log shown as:
opensmd.service - LSB: Manage OpenSM
Loaded: loaded (/etc/init.d/opensmd; generated)
Active: active (running) since Mon 2022-05-02 14:53:39 CST; 10s ago
#
# Warning: you may encounter the problem:
#
opensmd.service is not a native service, redirecting to systemd-sysv-install.
Executing: /lib/systemd/systemd-sysv-install enable opensmd
update-rc.d: error: no runlevel symlinks to modify, aborting!
#
# Please refer to the **Question #1** in FAQ for how to solve this problem
#Check the InfiniBand information
# Get the InfiniBand information
ibstat
# the log shown as:
# Adapter's stat should be Active.
Port 1:
State: Active
Physical state: LinkUp
Rate: 100
Base lid: 3
LMC: 0
SM lid: 3
Capability mask: 0x2651e84a
Port GUID: 0x0c42a10300605e88
Link layer: InfiniBandThe user needs to build remoteswap on both CPU server and memory servers.
Run rswap-server. This process must be alive all the time so either run it inside tmux or screen, or run it as a system service.
For now, you have to know the online core number of the CPU server first (sorry this hasn't been automated yet). You can check /proc/cpuinfo on the CPU server or simply get the number via top or htop.
A wrong core number will lead to crash.
./rswap-server <memory server ip> <memory server port> <memory pool size in GB> <number of cores on CPU server>
# an example: ./rswap-server 10.0.0.4 9400 48 32Edit the parameters in manage_rswap_client.sh.multi under $repo_home_dir/remoteswap/client directory.
Here is an excerpt of the script:
# The swap file/partition size should be equal to the whole size of remote memory
# Just example, edit them according to your machine
SWAP_PARTITION_SIZE="48"
server_ip="10.0.0.4"
server_port="9400"
swap_file="/mnt/swapfile"Make sure that "SWAP_PARTITION_SIZE" equal to the remote memory pool size you set when running rswap-server, as well as "server_ip" and "server_port" here.
To compile, enter $repo_home_dir/remoteswap/client on cpu server and $repo_home_dir/remoteswap/server on memory server, and just
makeTo install, on CPU server,
sudo ./manage_rswap_client.sh.multi installIt might take a while to allocate and register all memory in the memory pool, and establish the connection. The system should have been fully set up now.
You can optionally check system log via dmesg. A success should look like (1 chunk is 4GB so 12 chunks are essentially 48GB remote memory):
rswap_request_for_chunk, Got 12 chunks from memory server.
rdma_session_connect,Exit the main() function with built RDMA conenction rdma_session_context:0xffffffffc0e34a60 .
rswap_scheduler_init starts.
rswap_scheduler_init inits vqueues.
rswap_scheduler_init, wait for configuration to launch scheduler thd.
Swap RDMA bandwidth control functions registered.
frontswap module loaded
Setup done!
To uninstall, on CPU server,
sudo ./manage_rswap_client.sh.multi uninstallThe remoteswap kernel module exposes several tuning knobs for users to fine tune its performance. We offer a config helper and a default configuration. To update the config,
one can enter the $repo_home_dir/remoteswap/configure directory, and change the parameters in ProcConfig.txt and SchedulerConfig.txt. Then
sh apply.shFor more details, please refer to $repo_home_dir/remoteswap/README.md.
Currently Canvas uses cgroup-v1 to control application's memory limit. Here is an example.
Create cgroup on the CPU server:
# Create a memory cgroup, memctl, for the applications running on the CPU server.
# $USER is the username of the account. It's "guest" here.
sudo cgcreate -t $USER -a $USER -g memory:/memctl
# Please confirm that a memory cgroup directory is built as:
/sys/fs/cgroup/memory/memctlLimit local memory size on the CPU server. E.g., set local memory to 9GB:
# Please adjust the local memory size for different applications.
echo 9g > /sys/fs/cgroup/memory/memctl/memory.limit_in_bytesopensmd.service is not a native service, redirecting to systemd-sysv-install.
Executing: /lib/systemd/systemd-sysv-install enable opensmd
update-rc.d: error: no runlevel symlinks to modify, aborting!The original /etc/init.d/opensmd
8 ### BEGIN INIT INFO
9 # Provides: opensm
10 # Required-Start: $syslog openibd
11 # Required-Stop: $syslog openibd
12 # Default-Start: null
13 # Default-Stop: 0 1 6
14 # Description: Manage OpenSM
15 ### END INIT INFOChange the content in the line Default-start /etc/init.d/opensmd to :
12 # Default-Start: 2 3 4 5# Enable and strart the opensmd service
sudo update-rc.d opensmd remove -f
sudo systemctl enable opensmd
sudo systemctl start opensmd
# confirm the service status
sudo systemctl status opensmd
# The log shown as:
opensmd.service - LSB: Manage OpenSM
Loaded: loaded (/etc/init.d/opensmd; generated)
Active: active (running) since Mon 2022-05-02 14:53:39 CST; 10s ago
Thanks for all the constructive feedback that helped us with improving the performance and robustness of Canvas. We especially thanks Yulong Zhang, an undergraduate student from the University of Chinese Academy of Sciences (UCAS), who improved the efficiency and fairness of the two-dimensional RDMA scheduler. For additional questions please contact us at yifanqiao@g.ucla.edu or wangchenxi@ict.ac.cn.