Skip to content
/ gosbench Public

Distributed S3 benchmarking tool - Replacement of Cosbench

License

Notifications You must be signed in to change notification settings

mulbc/gosbench

Repository files navigation

Gosbench

Gosbench is the Golang reimplementation of Cosbench. It is a distributed S3 performance benchmark tool with Prometheus exporter leveraging the official Golang AWS SDK

Usage

Gosbench consists of two parts:

  • Server: Coordinates Workers and general test queue
  • Workers: Actually connect to S3 and perform reading, writing, deleting and listing of objects

INFO: -d activates debug logging, -t activates trace logging

Running a test

  1. Build the server: go install github.com/mulbc/gosbench/server
  2. Run the server, specifying a config file: server -c path/to/config.yaml - you can find an example config in the example folder
  3. The server will open port 2000 for workers to connect to - make sure this port is not blocked by your firewall!
  4. Build the worker: go install github.com/mulbc/gosbench/worker
  5. Run the worker, specifying the server connection details: worker -s 192.168.1.1:2000
  6. The worker will immediately connect to the server and will start to get to work. The worker opens port 8888 for the Prometheus exporter. Please make sure this port is allowed in your firewall and that you added the worker to the Prometheus config.

Prometheus configuration

Make sure your prometheus configuration looks similar to this:

global:
  scrape_interval:     15s # Set the scrape interval to every 15 seconds. Default is every 1 minute.
  evaluation_interval: 15s # Evaluate rules every 15 seconds. The default is every 1 minute.
scrape_configs:
  - job_name: 'prometheus'
    static_configs:
    - targets:
        - localhost:9090

  - job_name: 'gosbench'
    scrape_interval: 1s
    static_configs:
      - targets:
        - WORKER1.example.com:8888
        - WORKER2.example.com:8888

To reload the configuration, you can either send a SIGHUP to your prometheus server or just restart it ;) Afterwards ensure that you have your Gosbench workers listed at http://your.prometheus.server.example.com:9090/targets

It is expected that the workers are in state DOWN most of the time... they are only scrapeable during a test run.

It is Best Practice to run the Prometheus Node Exporter on all hosts as well, to gather common system metrics during the tests. This will help you in identifying bottlenecks. Please consult the Node Exporter manuals on how to install and configure it on your platform.

Evaluating a test

During a test, Prometheus will scrape the performance data continuously from the workers. You can visualize this data in Grafana. To get an overview of what the provided data looks like, check out the example scrape.

There is also an example Grafana dashboard that you can import and use. The Dashboard has some basic overview of the most common stats that people are interested in:

Gosbench Dashboard in action

Docker

There are now Docker container images available for easy consumption:

docker pull quay.io/mulbc/goroom-server
docker pull quay.io/mulbc/goroom-worker

In the k8s folder you will find example files to deploy Gosbench on Openshift and Kubernetes. Be sure to modify the ConfigMaps in gosbench.yaml to use your S3 endpoint credentials.

Reading pre-existing files from buckets

Due to popular demand, reading pre-existing files have been added. You activate this special mode by setting existing_read_weight to something higher than 0.

There are some important things to consider though ;)

Just like with other operations, the bucket_prefix value will be evaluated to determine the bucket name to search for pre-existing objects.

Example: This is an excerpt of your config:

    objects:
      size_min: 5
      size_max: 100
      part_size: 0
      # distribution: constant, random, sequential
      size_distribution: random
      unit: KB
      number_min: 10
      number_max: 100
      # distribution: constant, random, sequential
      number_distribution: constant
    buckets:
      number_min: 2
      number_max: 10
      # distribution: constant, random, sequential
      number_distribution: constant
    bucket_prefix: myBucket-

Note: Due to the constant distribution, we will only consider the _min values.

This will cause each workers to search for pre-existing files in the buckets myBucket-0 and myBucket-1 and read 10 objects from these buckets. If there are less than 10 objects in any of these buckets, some objects will be read multiple times. The object size given in your config will be ignored when reading pre-existing files.

Cosbench vs Gosbench benchmark comparision

When a new tool is presented, it’s essential to compare it to existing tools for accuracy. For this reason, we ran a comparision between Cosbench and Gosbench. Both benchmarks were tasked to do a 100% write test and 100% read test on 4KB, 16KB, 256KB, 1MB, 4MB objects for 60 seconds each. The tests were to run on one RGW using S3 protocol in ceph storage clusteri, also run in the test configuration in parallel. Figure below show writing and reading, respectively. From these charts, it’s apparent that the performance metrics for all objects are similar for both tools.

Latency

Bandwidth

Contributing

  • Be aware that this repo uses pre-commit hooks - install them via pre-commit install
  • We are using Go modules in this repository - read up on it here
  • Check out the open TODOs for hints on what to work on

Known issues

  • Workers will error out when the config's min value is larger than the max value (even for a constant distribution)