https://github.com/stackhpc/stackhpc-io-tools
Tools for running and post-processing fio benchmark data.
https://github.com/stackhpc/stackhpc-io-tools
Last synced: 14 days ago
JSON representation
Tools for running and post-processing fio benchmark data.
- Host: GitHub
- URL: https://github.com/stackhpc/stackhpc-io-tools
- Owner: stackhpc
- Created: 2018-10-25T15:15:14.000Z (over 7 years ago)
- Default Branch: master
- Last Pushed: 2023-09-05T09:44:37.000Z (almost 3 years ago)
- Last Synced: 2024-04-14T22:50:22.725Z (about 2 years ago)
- Language: Python
- Size: 6.59 MB
- Stars: 1
- Watchers: 7
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# To install
Installation from local files into a local virtualenv:
virtualenv venv
source venv/bin/activate
pip install .
Installation direct from the git repo:
pip install git+https://github.com/stackhpc/stackhpc-io-tools
# To build and push docker image
make docker DOCKER_ID=stackhpc
# To run fio locally (which launches a single client locally)
The `DATA_PATH` and `RESULTS_PATH` here refer to the path to read the
data from and dump results to locally.
make local SCENARIO=beegfs FIO_RW=randread NUM_CLIENTS=1 \
DATA_PATH=/path-to-test-dir RESULTS_PATH=/path-to-result-dir
# To run remote baremetal batch jobs:
This expects you to be able to SSH into the nodes normally and `fio_jobfiles` is available in the home directory of the default user which you can copy over using the following:
make copy NUM_NODES=2 NODE_PREFIX=centos@kata-worker
Then to invoke the batch job:
for c in 1 64; do
for rw in write randwrite read randread; do
make remote SCENARIO=beegfs FIO_RW=$rw NUM_CLIENTS=$c \
DATA_PATH=/mnt/storage-nvme \
RESULTS_PATH=/mnt/storage-nvme/bharat/results/bare \
NUM_NODES=2 NODE_PREFIX=centos@kata-worker
done
done
If you have a large number of clients, make sure that you update the
MaxStartups paramater in `/etc/ssh/sshd_config` and restart `sshd` in worker
nodes otherwise your connections may get dropped because the default setting of
`10:30:50` means that `30%` of connections when there are more than `10` will get
dropped up to a maximum of `50` after which all connections are dropped.
# To run as a Kubernetes job:
The `DATA_HOSTPATH` and `RESULTS_HOSTPATH` here refer to the path to read the
data from and dump results to on the Kubernetes worker nodes. These paths are
mapped to the default `DATA_PATH` and `RESULTS_PATH` inside containers spawned
by Kubernetes.
for c in 1 64; do
for rw in write randwrite read randread; do
make k8s SCENARIO=beegfs FIO_RW=$rw NUM_CLIENTS=$c \
DATA_HOSTPATH=/mnt/storage-nvme/bharat \
RESULTS_HOSTPATH=/mnt/storage-nvme/bharat/results/runc
done
done
There is an option to add `RUNTIME_CLASS` if it is supported by the Kubernetes cluster:
for c in 1 64; do
for rw in write randwrite read randread; do
make k8s SCENARIO=beegfs FIO_RW=$rw NUM_CLIENTS=$c \
DATA_HOSTPATH=/mnt/storage-nvme \
RESULTS_HOSTPATH=/mnt/storage-nvme/bharat/results/kata \
RUNTIME_CLASS=kata-qemu
done
done
As the test is running, it might be useful to look at the verbose output of the test as follows:
kubectl --namespace default logs jobs/beegfs-randread-1 --follow
For debugging, you can also invoke shell inside the pod:
kubectl --namespace default exec -it beegfs-randread-1-n44kj sh
# To generate plot:
make parse RESULTS_PATH=/path-to-result-dir OUTPUT_PATH=/path-to-output-dir
NOTE: `RESULTS_PATH` is the path for raw data generated by `fio` and
`OUTPUT_PATH` is the parsed data analysed by `bin/fio_parse` executable
presented in this repo.
# Typical output figures:
