Tebis is a persistent LSM key value (KV) store desinged for fast storage devices and RDMA networks. Tebis uses two main technologies :
- Hybrid KV placement via its Parallax LSM KV store to reduce I/O amplification and increase CPU efficiency.
- Index shipping via CPU efficient RDMA bulk network transfers to reduce the compaction overhead in replicas. Instead of repeating the compaction in replicas primary ships the index which replicas rewrite to be valid for their storage address space.
More details can be found in the Eurosys '22 paper Tebis: Index Shipping for Efficient Replication in LSM Key-Value Stores.
- YCSB-CXX contains the C++ version of the YCSB benchmark along with a Tebis driver
- tebis_rdma contains code rdma utilities used in the project
- tebis_server contains all the server related code
- tebis_rdma_client contains the client side code of Tebis
- File tebis_rdma_client/tebis_rdma_client.h contains the public API of the client API
- tcp_server contains the code for a standalone tcp_server over Parallax, it is a separate cmake project. Detailed information are in the tcp_server folder README.md
- tcp_client contains the code for the TCP client
Note: It has been tested with gcc 10.1.0
To build Tebis, the following libraries have to be installed on your system:
libnuma
- Allocations with NUMA policylibibverbs
- Infiniband verbslibrdmacm
- RDMA Connection Managerlibzookeeper_mt
- Zookeeper client bindings for C
For Mellanox cards, the Infiniband and RDMA libraries are included in the software package provided by the vendor. Additionally, Tebis uses cmake for its build system and the gcc and g++ compilers for its compilation.
Tebis requires CMake version >= 3.11.0. On Ubuntu, you need to add the following repository to get the latest stable version of CMake:
wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | sudo apt-key add -
sudo apt-add-repository 'deb https://apt.kitware.com/ubuntu/ bionic main'
sudo apt update
Run the following command with superuser privileges:
sudo apt install libnuma-dev libibverbs-dev librdmacm-dev libzookeeper-mt-dev
For the build tools and compiler:
sudo apt install cmake build-essential
Tebis requires CMake version >= 3.11.0. On Centos/RHEL this is supplied from the EPEL repository and can be installed with:
sudo yum install cmake3 kernel-devel gcc-c++
sudo yum install numactl-devel boost-devel
For RDMA:
sudo yum install libibverbs-devel librdmacm-devel
You also need to install ZooKeeper. Ready-made packages are available from Cloudera.
-
Add Cloudera's Centos 7 repository as described here
-
Install the Zookeeper C binding for clients:
yum install zookeeper-native
Compilation is done using the clang compiler, provided by the clang package in most Linux distributions. To configure the build system of Tebis and build it run the commands:
mkdir build
cd build
cmake .. -DTEBIS_FORMAT=ON
make
On Centos/RHEL 7, replace the cmake
command with the cmake3
command supplied
from the EPEL package of the same name.
In the case of the standalone tcp-server you need to allocate a file for Parallax via the command
fallocate -l <size in GB>G <file name>
Then, you need to initialize it with the kv_format tool of Parallax
<BUILD_FOLDER>_deps/parallax-build/lib/kv_format.parallax --device <file name> --max_regions_num <number of regions>
Finally, execute
<BUILD_FOLDER>/tcp_server/tcp-server -t <threads num> -b <IP address> -p <port number> -f <file name> -L0 <L0 size in MB> -GF <growth factor>
The CMake scripts provided support two build configurations; "Release" and "Debug". The Debug configuration enables the "-g" option during compilation to allow debugging. The build configuration can be defined as a parameter to the cmake call as follows:
cmake3 .. -DCMAKE_BUILD_TYPE="Debug|Release"
The default build configuration is "Debug".
The "Release" build disables warnings and enables optimizations.
- build/tebis_server/libtebis_client.a - Client library for applications to interact with Tebis
- build/tebis_server/tebis_server - The executable of Tebis server
- build/YCSB-CXXX/ycsb-async-tebis - YCSB that uses Tebis as its storage
Install the clang static analyzer with the command:
sudo pip install scan-build
Before running the analyzer, make sure to delete any object files and executables from previous build by running in the root of the repository:
rm -r build
Then generate a report using:
scan-build --intercept-first make
The last line of the above command's output will mention the folder where the newly created report resides in. For example:
"scan-build: Run 'scan-view /tmp/scan-build-2018-09-05-16-21-31-978968-9HK0UO'
to examine bug reports."
To view the report you can run the above command, assuming you have a graphical environment or just copy the folder mentioned to a computer that does and open the index.html file in that folder.
To install pre-commit:
pip install pre-commit --user
pre-commit --version
2.2.0
If the above command does not print 2.2.0 you need to update python using:
sudo yum update python3
Then try upgrading pre-commit:
pip install -U pre-commit --user
To install pre-commit hooks:
cd tebis
pre-commit install
pre-commit install --hook-type commit-msg
If everything worked as it should then the following message should be printed: pre-commit installed at .git/hooks/pre-commit
If you want to run a specific hook with a specific file run:
pre-commit run hook-id --files filename
pre-commit run cmake-format --files CMakeLists.txt
Tebis uses RDMA for all network communication, which requires support from the network interface to run. A software implementation (soft-RoCE) exists and can run on all network interfaces.
soft-RoCE is part of the mainline Linux kernel versions since version 4.9
through the rdma_rxe
kernel module. To enable it for a network adapter the
following steps are required:
The ibverbs-utils
and rdma-core
packages are required to enable soft-RoCE.
These packages should be in most distirbutions' repositories
Run the following command form a terminal with root access (or use sudo):
# apt install ibverbs-utils rdma-core perftest
To enable soft-RoCE on a network command run the following commands with superuser privileges:
rxe_cfg start
rxe_cfg add eth_interface
where eth_interface
is the name of an ethernet network adapter interface. To
view available network adapters run ip a
.
The command rxe_cfg start
has to be run at every boot to use RDMA features.
To verify that soft-RoCE is working, we can run a simple RDMA Write throuhgput benchmark.
First, open two shells, one to act as the server and one to act as the client. Then run the following commands:
- On the server:
ib_write_bw
- On the client:
ib_write_bw eth_interface_ip
, whereeth_interface_ip
is the IP address of a soft-RoCE enabled ethernet interface.
Example output:
- Server process:
************************************
* Waiting for client to connect... *
************************************
---------------------------------------------------------------------------------------
RDMA_Write BW Test
Dual-port : OFF Device : rxe0
Number of qps : 1 Transport type : IB
Connection type : RC Using SRQ : OFF
CQ Moderation : 100
Mtu : 1024[B]
Link type : Ethernet
GID index : 1
Max inline data : 0[B]
rdma_cm QPs : OFF
Data ex. method : Ethernet
---------------------------------------------------------------------------------------
local address: LID 0000 QPN 0x0011 PSN 0x3341fd RKey 0x000204 VAddr 0x007f7e1b8fa000
GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:122:205
remote address: LID 0000 QPN 0x0012 PSN 0xbfbac5 RKey 0x000308 VAddr 0x007f70f5843000
GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:122:205
---------------------------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
65536 5000 847.44 827.84 0.013245
---------------------------------------------------------------------------------------
- Client process:
---------------------------------------------------------------------------------------
RDMA_Write BW Test
Dual-port : OFF Device : rxe0
Number of qps : 1 Transport type : IB
Connection type : RC Using SRQ : OFF
TX depth : 128
CQ Moderation : 100
Mtu : 1024[B]
Link type : Ethernet
GID index : 1
Max inline data : 0[B]
rdma_cm QPs : OFF
Data ex. method : Ethernet
---------------------------------------------------------------------------------------
local address: LID 0000 QPN 0x0012 PSN 0xbfbac5 RKey 0x000308 VAddr 0x007f70f5843000
GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:122:205
remote address: LID 0000 QPN 0x0011 PSN 0x3341fd RKey 0x000204 VAddr 0x007f7e1b8fa000
GID: 00:00:00:00:00:00:00:00:00:00:255:255:192:168:122:205
---------------------------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
65536 5000 847.44 827.84 0.013245
---------------------------------------------------------------------------------------
git config commit.template .git-commit-template
You can use cgroups
to limit the memory available to a process by running the
process using systemd-run
. In addition to memory allocations, pages in the Linux
kernel's buffer cache count towards the cgroups
memory limit. Example usage:
# systemd-run --unit=unit0 --scope --slice=slice0 --property MemoryLimit=16G <command>
Starting a systemd
unit requires root privileges. The above example will limit
the memory available to a command (including pages in the buffer cache) to 16GB.
First we need a Zookeeper server. For simplicity we assume that the Zookeeper service runs at zoo:2181. Then we need to initialize Tebis metadata. This can be done through the command <tebis_root_folder>/scripts/tebis/tebis_zk_init.py <hosts_file> <regions_file> <zookeeper_host>
Hosts_file: Contains the servers of the cluster in the form host1:port_for_incoming_rdma_connections:0
Example:
sith2.cluster.ics.forth.gr:8080:0 leader #(so sith2.cluster.ics.forth.gr:8080 will be the initial leader of the system)
sith3.cluster.ics.forth.gr:8080:0
sith6.cluster.ics.forth.gr:8080:0
Regions file: Contains the region info in which we split the key space
<region_id> <min_key_range> <max_key_range> <server1:port:1 (primary)> <server2:port:1 (backup)>
Example:
0 -oo MM sith2.cluster.ics.forth.gr:8080:1 sith3.cluster.ics.forth.gr:8080:1
1 MM ZZ sith3.cluster.ics.forth.gr:8080:1 sith6.cluster.ics.forth.gr:8080:1
2 ZZ +oo sith6.cluster.ics.forth.gr:8080:1 sith2.cluster.ics.forth.gr:8080:1
In each tebis server we need a allocated file where Tebis will store its data. Each server's storage capacity will be equal
to the size of the file provided. The server will create its own file (or use dd or fallocate).
Example: fallocate -l 100G /path/to/file
Then we need to boot first the leader of the Tebis rack
<tebis_build_root folder>/tebis_server/tebis_server -d <path to tebis file> -z <zk_host:zk_port> -r <RDMA IP subnet> -p <server port> -c <num of threads> [-t <LSM L0 size in keys>] [-g <growth factor>] [-i <"send_index" | "build_index">] [-s <device size in GB>]
example: build/tebis_server/tebis_server -d /nvme/par1.dat -z sith2:2181 -r 192.168.4 -p 8080 -c 3
example: build/tebis_server/tebis_server -d /nvme/par1.dat -z sith2:2181 -r 192.168.4 -p 8080 -c 3 -t 16 -g 10 -i send_index -s 100
cd into folder <BUILD_ROOT_FOLDER>/tests/
and type
test_krc_api zk_host:zk_port
We thankfully acknowledge the support of the European Commission under the Horizon 2020 Framework Programme for Research and Innovation through the projects EVOLVE (Grant Agreement ID: 825061). This work is (also) partly supported by project EUPEX, which has received funding from the European High-Performance Computing Joint Undertaking (JU) under grant agreement No 101033975. The JU receives support from the European Union's Horizon 2020 re-search and innovation programme and France, Germany, Italy, Greece, United Kingdom, Czech Republic, Croatia.