DINOMO: An Elastic, Scalable, High-Performance Key-Value Store for Disaggregated Persistent Memory (VLDB 2023, PVLDB 2022 Vol. 15 No. 13)
Dinomo is a novel key-value store for disaggregated persistent memory (DPM). Dinomo is the first key-value store for DPM that simultaneously achieves high common-case performance, scalability, and lightweight online reconfiguration. This repository includes the implementations of our key-value store and relevant tools to run experiments.
Please cite the following paper if you use Dinomo or compare it with your system:
@article{dinomo-vldb,
author = {Lee, Sekwon and Ponnapalli, Soujanya and Singhal, Sharad and Aguilera, Marcos K. and Keeton, Kimberly and Chidambaram, Vijay},
title = {DINOMO: An Elastic, Scalable, High-Performance Key-Value Store for Disaggregated Persistent Memory},
year = {2022},
publisher = {VLDB Endowment},
volume = {15},
number = {13},
issn = {2150-8097},
url = {https://doi.org/10.14778/3565838.3565854},
doi = {10.14778/3565838.3565854},
journal = {Proc. VLDB Endow.},
pages = {4023 - 4037},
numpages = {15}
}
include/dinomo_compute.hpp
: Core implementation ofKN
.src/kvs/adaptive-cache.h
: Core implementation ofDAC
.src/kvs/dinomo_storage.cpp
: Core implementation ofDPM
.src/monitor/monitoring.cpp
: Core implementation ofM-node
.conf/dinomo-base.yml
: Configuration file for KVS
Dinomo uses Kubernetes as a cluster orchestration tool.
We will build container images for each instance pod and make Kubernetes
automatically manage them. Dinomo cluster consists of a Monitoring/management
node (M-node
), Routing Nodes (RNs
), KVS Nodes (KNs
), DPM
, and
clients (running inside applications). To run Dinomo, the current implementation
needs at least 5 separate physical machines to assign each pod for the node
instances exclusively to a separate physical machine. Among those physical machines,
at least 2 machines must be Infiniband-enabled for a KN
and a DPM
.
We use Kubespray to generate
Dinomo cluster orchestrated by Kubernetes over baremetal servers.
In the following instructions, we assume that a setting where 1 M-node
, 1 RN
, 1 KN
,
1 DPM
, and 1 application client exist. However, you can also increase the number
of RNs
and KNs
according to your own setting. We assume the instructions
specified in the following sections are excuted in one of the baremetal servers.
This server will act as a master node where a Kubernetes master as well as a
M-node
pod are spawned and as a domain to control the overall cluster executions
with a command line interface.
- Ubuntu18.04
- Connectx-3 RNIC
$ cd ~/; mkdir projects; cd projects; git clone https://github.com/utsaslab/dinomo.git DINOMO; cd DINOMO
$ bash scripts/dependencies.sh
There are a few environment variables that should be configured according to your own environment.
$DINOMO_HOME
: Path to Dinomo home directory in the master node$REMOTE_USER_NAME
: Other machines's user name, user name should be same across all the machines since the python script to generate cluster uses the user name to access other machines over ssh and install required packages remotely.
$ export DINOMO_HOME=/home/cc/projects/DINOMO/
$ export REMOTE_USER_NAME=cc
$ ulimit -n $(ulimit -Hn)
Update configuration files by specifying the ip addresses of nodes depending on their type.
$ vi kubespray/inventory/dinomo_cluster/inventory.ini
$ vi cluster/dinomo/cluster/kubespray/benchmarkNodes
$ vi cluster/dinomo/cluster/kubespray/masterNode
$ vi cluster/dinomo/cluster/kubespray/memoryNodes
$ vi cluster/dinomo/cluster/kubespray/routingNodes
$ vi cluster/dinomo/cluster/kubespray/storageNodes
- Run the following python script to build up cluster configurations over baremetal servers. The script automatically accesses each server over ssh to apply the required configures.
Usage: python3 -m dinomo.cluster.create_cluster -m <# of KNs> -r <# of RNs> -b <# of ClientNodes>
$ cd cluster
$ python3 -m dinomo.cluster.create_cluster -m 1 -r 1 -b 1
- Check if the cluster configurations are completed properly. The following command should show the list of the server nodes currently managed by Kubernetes.
$ kubectl get nodes
$ cd ~/projects/DINOMO
$ bash scripts/build.sh -bRelease -j8 -g
We provide the pre-built container images for Dinomo cluster through our public repository. If you want to use your own images, please refer to the following script.
$ bash scripts/clean_build_docker.sh
We execute Dinomo cluster on the master node and one of the Infiniband-enabled servers.
Note that the python script we used will only set up environments (installing required packages) and run nodes as Kubernetes pods except for DPM node. So, you should run DPM instance manually in the DPM machine you will use. Please download this source code to DPM side as well and follow the instructions below to run.
$ cd ~/projects/DINOMO
$ bash scripts/dependencies.sh
$ bash scripts/build.sh -bRelease -j8 -g
$ sudo ./build/target/kvs/dinomo-storage
Usage: python3 -m dinomo.cluster.run_cluster -m <# of KNs> -r <# of RNs> -b <# of ClientNodes>
$ cd cluster
$ python3 -m dinomo.cluster.run_cluster -m 1 -r 1 -b 1
After generating pods, the script will prompt an event trigger. The trigger supports seven types of commands to allow us to manually control the cluster. We used this trigger to manually control the load variance in our evaluation section.
-
LOAD
: load the given number of key-value pairs to KVS (This command must be first run before running workloads). -
RUN
: run a workload with particular read-write ratio and key distribution. -
WARM
: warm up KVS with the given number of requests as well as the specified workload pattern.
WARM-RUN
: run the workload again used in warm-up period for a specified time. This command should be used coupled withWARM
.
-
ADD
: spawn more application nodes (clients). -
REMOVE
: remove the specified number of application nodes (clients). -
FAIL
: simulate failures to the specified number of KNs.
Please contact us at sklee@cs.utexas.edu
or vijayc@utexas.edu
with any questions.
Some part of DINOMO's codebase from Anna.