PerfKit Benchmarker is an open effort to define a canonical set of benchmarks to measure and compare cloud offerings. It's designed to operate via vendor provided command line tools. The benchmark default settings are not tuned for any particular platform or instance type. These settings are recommended for consistency across services. Only in the rare case where there is a common practice like setting the buffer pool size of a database do we change any settings.

This README is designed to give you the information you need to get running with the benchmarker and the basics of working with the code. The wiki contains more detailed information:

• FAQ
• Tech Talks
• Governing rules
• Community meeting decks and notes
• Design documents
• You are always welcome to open an issue, or to join us on #PerfKitBenchmarker on freenode to discuss issues you're having, pull requests, or anything else related to PerfKitBenchmarker

PerfKit Benchmarker provides wrappers and workload definitions around popular benchmark tools. We made it very simple to use and automate everything we can. It instantiates VMs on the Cloud provider of your choice, automatically installs benchmarks, and runs the workloads without user interaction.

Due to the level of automation you will not see prompts for software installed as part of a benchmark run. Therefore you must accept the license of each of the benchmarks individually, and take responsibility for using them before you use the PerfKit Benchmarker.

Moving forward, you will need to run PKB with the explicit flag --accept-licenses.

In its current release these are the benchmarks that are executed:

• aerospike: Apache v2 for the client and GNU AGPL v3.0 for the server
• bonnie++: GPL v2
• cassandra_ycsb: Apache v2
• cassandra_stress: Apache v2
• cloudsuite3.0: CloudSuite 3.0 license
• cluster_boot: MIT License
• coremark: EEMBC
• copy_throughput: Apache v2
• fio: GPL v2
• gpu_pcie_bandwidth: NVIDIA Software Licence Agreement
• hadoop_terasort: Apache v2
• hpcc: Original BSD license
• hpcg: BSD 3-clause
• iperf: UIUC License
• memtier_benchmark: GPL v2
• mesh_network: HP license
• mongodb: Deprecated. GNU AGPL v3.0
• mongodb_ycsb: GNU AGPL v3.0
• multichase: Apache v2
• netperf: HP license
• oldisim: Apache v2
• object_storage_service: Apache v2
• pgbench: PostgreSQL Licence
• ping: No license needed.
• silo: MIT License
• scimark2: public domain
• speccpu2006: SPEC CPU2006
• SHOC: BSD 3-clause
• sysbench_oltp: GPL v2
• TensorFlow: Apache v2
• tomcat: Apache v2
• unixbench: GPL v2
• wrk: Modified Apache v2
• ycsb (used by mongodb, hbase_ycsb, and others): Apache v2

Some of the benchmarks invoked require Java. You must also agree with the following license:

• openjdk-7-jre: GPL v2 with the Classpath Exception

SPEC CPU2006 benchmark setup cannot be automated. SPEC requires that users purchase a license and agree with their terms and conditions. PerfKit Benchmarker users must manually download cpu2006-1.2.iso from the SPEC website, save it under the perfkitbenchmarker/data folder (e.g. ~/PerfKitBenchmarker/perfkitbenchmarker/data/cpu2006-1.2.iso), and also supply a runspec cfg file (e.g. ~/PerfKitBenchmarker/perfkitbenchmarker/data/linux64-x64-gcc47.cfg). Alternately, PerfKit Benchmarker can accept a tar file that can be generated with the following steps:

• Extract the contents of cpu2006-1.2.iso into a directory named cpu2006
• Run cpu2006/install.sh
• Copy the cfg file into cpu2006/config
• Create a tar file containing the cpu2006 directory, and place it under the perfkitbenchmarker/data folder (e.g. ~/PerfKitBenchmarker/perfkitbenchmarker/data/cpu2006v1.2.tgz).

PerfKit Benchmarker will use the tar file if it is present. Otherwise, it will search for the iso and cfg files.

Before you can run the PerfKit Benchmarker, you need account(s) on the cloud provider(s) you want to benchmark (see providers). You also need the software dependencies, which are mostly command line tools and credentials to access your accounts without a password. The following steps should help you get up and running with PKB.

The recommended way to install and run PKB is in a virtualenv with the latest version of Python 3 (at least Python 3.9). Most Linux distributions and recent Mac OS X versions already have Python 3 installed at /usr/bin/python3.

If Python is not installed, you can likely install it using your distribution's package manager, or see the Python Download page.

python3 -m venv $HOME/my_virtualenv
source $HOME/my_virtualenv/bin/activate

Download the latest PerfKit Benchmarker release from GitHub. You can also clone the working version with:

$ cd $HOME
$ git clone https://github.com/GoogleCloudPlatform/PerfKitBenchmarker.git

Install Python library dependencies:

$ pip3 install -r $HOME/PerfKitBenchmarker/requirements.txt

You may need to install additional dependencies depending on the cloud provider you are using. For example, for AWS:

$ cd $HOME/PerfKitBenchmarker/perfkitbenchmarker/providers/aws
$ pip3 install -r requirements.txt

Some benchmarks may require data to be preprovisioned in a cloud. To preprovision data, you will need to obtain the data and then upload it to that cloud. See more information below about which benchmarks require preprovisioned data and how to upload it to different clouds.

PerfKit Benchmarker can run benchmarks both on Cloud Providers (GCP, AWS, Azure, DigitalOcean) as well as any "machine" you can SSH into.

$ ./pkb.py --project=<GCP project ID> --benchmarks=iperf --machine_type=f1-micro

$ cd PerfKitBenchmarker
$ ./pkb.py --cloud=AWS --benchmarks=iperf --machine_type=t2.micro

$ ./pkb.py --cloud=Azure --machine_type=Standard_A0 --benchmarks=iperf

$ ./pkb.py --cloud=IBMCloud --machine_type=cx2-4x8 --benchmarks=iperf

$ ./pkb.py --cloud=AliCloud --machine_type=ecs.s2.large --benchmarks=iperf

$ ./pkb.py --cloud=DigitalOcean --machine_type=16gb --benchmarks=iperf

$ ./pkb.py --cloud=OpenStack --machine_type=m1.medium \
           --openstack_network=private --benchmarks=iperf

$ ./pkb.py --cloud=Kubernetes --benchmarks=iperf --kubectl=/path/to/kubectl --kubeconfig=/path/to/kubeconfig --image=image-with-ssh-server  --ceph_monitors=10.20.30.40:6789,10.20.30.41:6789

$ ./pkb.py --cloud=Mesos --benchmarks=iperf --marathon_address=localhost:8080 --image=image-with-ssh-server

./pkb.py --cloud=CloudStack --benchmarks=ping --cs_network_offering=DefaultNetworkOffering

$ ./pkb.py --cloud=Rackspace --machine_type=general1-2 --benchmarks=iperf

$ ./pkb.py --cloud=ProfitBricks --machine_type=Small --benchmarks=iperf

Install all dependencies as above and ensure that smbclient is installed on your system if you are running on a linux controller:

$ which smbclient
/usr/bin/smbclient

Now you can run Windows benchmarks by running with --os_type=windows. Windows has a different set of benchmarks than Linux does. They can be found under perfkitbenchmarker/windows_benchmarks/. The target VM OS is Windows Server 2012 R2.

Juju is a service orchestration tool that enables you to quickly model, configure, deploy and manage entire cloud environments. Supported benchmarks will deploy a Juju-modeled service automatically, with no extra user configuration required, by specifying the --os_type=juju flag.

$ ./pkb.py --cloud=AWS --os_type=juju --benchmarks=cassandra_stress

Benchmark/Package authors need to implement the JujuInstall() method inside their package. This method deploys, configures, and relates the services to be benchmarked. Please note that other software installation and configuration should be bypassed when FLAGS.os_type == JUJU. See perfkitbenchmarker/linux_packages/cassandra.py for an example implementation.

Run with --benchmarks="standard_set" and every benchmark in the standard set will run serially which can take a couple of hours. Additionally, if you don't specify --cloud=..., all benchmarks will run on the Google Cloud Platform.

Named sets are are groupings of one or more benchmarks in the benchmarking directory. This feature allows parallel innovation of what is important to measure in the Cloud, and is defined by the set owner. For example the GoogleSet is maintained by Google, whereas the StanfordSet is managed by Stanford. Once a quarter a meeting is held to review all the sets to determine what benchmarks should be promoted to the standard_set. The Standard Set is also reviewed to see if anything should be removed. To run all benchmarks in a named set, specify the set name in the benchmarks parameter (e.g., --benchmarks="standard_set"). Sets can be combined with individual benchmarks or other named sets.

The following are some common flags used when configuring PerfKit Benchmarker.

The default cloud is 'GCP', override with the --cloud flag. Each cloud has a default zone which you can override with the --zones flag, the flag supports the same values that the corresponding Cloud CLIs take:

Example:

./pkb.py --cloud=GCP --zones=us-central1-a --benchmarks=iperf,ping

The disk type names vary by provider, but the following table summarizes some useful ones. (Many cloud providers have more disk types beyond these options.)

Also note that --data_disk_type=local tells PKB not to allocate a separate disk, but to use whatever comes with the VM. This is useful with AWS instance types that come with local SSDs, or with the GCP --gce_num_local_ssds flag.

If an instance type comes with more than one disk, PKB uses whichever does not hold the root partition. Specifically, on Azure, PKB always uses /dev/sdb as its scratch device.

If the VM guests do not have direct Internet access in the cloud environment, you can configure proxy settings through pkb.py flags.

To do that simple setup three flags (All urls are in notation ): The flag values use the same <protocol>://<server>:<port> syntax as the corresponding environment variables, for example --http_proxy=http://proxy.example.com:8080 .

As mentioned above, some benchmarks require preprovisioned data. This section describes how to preprovision this data.

This benchmark demonstrates the use of preprovisioned data. Create the following file to upload using the command line:

echo "1234567890" > preprovisioned_data.txt

To upload, follow the instructions below with a filename of preprovisioned_data.txt and a benchmark name of sample.

To preprovision data on Google Cloud, you will need to upload each file to Google Cloud Storage using gsutil. First, you will need to create a storage bucket that is accessible from VMs created in Google Cloud by PKB. Then copy each file to this bucket using the command

gsutil cp <filename> gs://<bucket>/<benchmark-name>/<filename>

To run a benchmark on Google Cloud that uses the preprovisioned data, use the flag --gcp_preprovisioned_data_bucket=<bucket>.

To preprovision data on AWS, you will need to upload each file to S3 using the AWS CLI. First, you will need to create a storage bucket that is accessible from VMs created in AWS by PKB. Then copy each file to this bucket using the command

aws s3 cp <filename> s3://<bucket>/<benchmark-name>/<filename>

To run a benchmark on AWS that uses the preprovisioned data, use the flag --aws_preprovisioned_data_bucket=<bucket>.

Each benchmark now has an independent configuration which is written in YAML. Users may override this default configuration by providing a configuration. This allows for much more complex setups than previously possible, including running benchmarks across clouds.

A benchmark configuration has a somewhat simple structure. It is essentially just a series of nested dictionaries. At the top level, it contains VM groups. VM groups are logical groups of homogenous machines. The VM groups hold both a vm_spec and a disk_spec which contain the parameters needed to create members of that group. Here is an example of an expanded configuration:

hbase_ycsb:
  vm_groups:
    loaders:
      vm_count: 4
      vm_spec:
        GCP:
          machine_type: n1-standard-1
          image: ubuntu-16-04
          zone: us-central1-c
        AWS:
          machine_type: m3.medium
          image: ami-######
          zone: us-east-1a
        # Other clouds here...
      # This specifies the cloud to use for the group. This allows for
      # benchmark configurations that span clouds.
      cloud: AWS
      # No disk_spec here since these are loaders.
    master:
      vm_count: 1
      cloud: GCP
      vm_spec:
        GCP:
          machine_type:
            cpus: 2
            memory: 10.0GiB
          image: ubuntu-16-04
          zone: us-central1-c
        # Other clouds here...
      disk_count: 1
      disk_spec:
        GCP:
          disk_size: 100
          disk_type: standard
          mount_point: /scratch
        # Other clouds here...
    workers:
      vm_count: 4
      cloud: GCP
      vm_spec:
        GCP:
          machine_type: n1-standard-4
          image: ubuntu-16-04
          zone: us-central1-c
        # Other clouds here...
      disk_count: 1
      disk_spec:
        GCP:
          disk_size: 500
          disk_type: remote_ssd
          mount_point: /scratch
        # Other clouds here...

For a complete list of keys for vm_specs and disk_specs see virtual_machine.BaseVmSpec and disk.BaseDiskSpec and their derived classes.

User configs are applied on top of the existing default config and can be specified in two ways. The first is by supplying a config file via the --benchmark_config_file flag. The second is by specifying a single setting to override via the --config_override flag.

A user config file only needs to specify the settings which it is intended to override. For example if the only thing you want to do is change the number of VMs for the cluster_boot benchmark, this config is sufficient:

cluster_boot:
  vm_groups:
    default:
      vm_count: 100

You can achieve the same effect by specifying the --config_override flag. The value of the flag should be a path within the YAML (with keys delimited by periods), an equals sign, and finally the new value:

--config_override=cluster_boot.vm_groups.default.vm_count=100

See the section below for how to use static (i.e. pre-provisioned) machines in your config.

It is possible to run PerfKit Benchmarker without running the Cloud provisioning steps. This is useful if you want to run on a local machine, or have a benchmark like iperf run from an external point to a Cloud VM.

In order to do this you need to make sure:

• The static (i.e. not provisioned by PerfKit Benchmarker) machine is ssh'able
• The user PerfKitBenchmarker will login as has 'sudo' access. (*** Note we hope to remove this restriction soon ***)

Next, you will want to create a YAML user config file describing how to connect to the machine as follows:

static_vms:
  - &vm1 # Using the & character creates an anchor that we can
         # reference later by using the same name and a * character.
    ip_address: 170.200.60.23
    user_name: voellm
    ssh_private_key: /home/voellm/perfkitkeys/my_key_file.pem
    zone: Siberia
    disk_specs:
      - mount_point: /data_dir

• The ip_address is the address where you want benchmarks to run.
• ssh_private_key is where to find the private ssh key.
• zone can be anything you want. It is used when publishing results.
• disk_specs is used by all benchmarks which use disk (i.e., fio, bonnie++, many others).

In the same file, configure any number of benchmarks (in this case just iperf), and reference the static VM as follows:

iperf:
  vm_groups:
    vm_1:
      static_vms:
        - *vm1

I called my file iperf.yaml and used it to run iperf from Siberia to a GCP VM in us-central1-f as follows:

$ ./pkb.py --benchmarks=iperf --machine_type=f1-micro --benchmark_config_file=iperf.yaml --zones=us-central1-f --ip_addresses=EXTERNAL

• ip_addresses=EXTERNAL tells PerfKit Benchmarker not to use 10.X (ie Internal) machine addresses that all Cloud VMs have. Just use the external IP address.

If a benchmark requires two machines like iperf, you can have two machines in the same YAML file as shown below. This means you can indeed run between two machines and never provision any VMs in the Cloud.

static_vms:
  - &vm1
    ip_address: <ip1>
    user_name: connormccoy
    ssh_private_key: /home/connormccoy/.ssh/google_compute_engine
    internal_ip: 10.240.223.37
    install_packages: false
  - &vm2
    ip_address: <ip2>
    user_name: connormccoy
    ssh_private_key: /home/connormccoy/.ssh/google_compute_engine
    internal_ip: 10.240.234.189
    ssh_port: 2222

iperf:
  vm_groups:
    vm_1:
      static_vms:
        - *vm2
    vm_2:
      static_vms:
        - *vm1

You can now specify flags in configuration files by using the flags key at the top level in a benchmark config. The expected value is a dictionary mapping flag names to their new default values. The flags are only defaults; it's still possible to override them via the command line. It's even possible to specify conflicting values of the same flag in different benchmarks:

iperf:
  flags:
    machine_type: n1-standard-2
    zone: us-central1-b
    iperf_sending_thread_count: 2

netperf:
  flags:
    machine_type: n1-standard-8

The new defaults will only apply to the benchmark in which they are specified.

PerfKit data can optionally be published to an Elasticsearch server. To enable this, the elasticsearch Python package must be installed.

$ pip install elasticsearch

Note: The elasticsearch Python library and Elasticsearch must have matching major versions.

The following are flags used by the Elasticsearch publisher. At minimum, all that is needed is the --es_uri flag.

Note: Amazon ElasticSearch service currently does not support transport on port 9200 therefore you must use endpoint with port 80 eg. search-<ID>.es.amazonaws.com:80 and allow your IP address in the cluster.

No additional packages need to be installed in order to publish Perfkit data to an InfluxDB server.

InfluxDB Publisher takes in the flags for the Influx uri and the Influx DB name. The publisher will default to the pre-set defaults, identified below, if no uri or DB name is set. However, the user is required to at the very least call the --influx_uri flag to publish data to Influx.

First start with the CONTRIBUTING.md file. It has the basics on how to work with PerfKitBenchmarker, and how to submit your pull requests.

In addition to the CONTRIBUTING.md file we have added a lot of comments into the code to make it easy to:

• Add new benchmarks (e.g.: --benchmarks=<new benchmark>)
• Add new package/os type support (e.g.: --os_type=<new os type>)
• Add new providers (e.g.: --cloud=<new provider>)
• etc.

Even with lots of comments we make to support more detailed documention. You will find the documentation we have on the wiki. Missing documentation you want? Start a page and/or open an issue to get it added.

If you wish to run unit or integration tests, ensure that you have tox >= 2.0.0 installed.

In addition to regular unit tests, which are run via hooks/check-everything, PerfKit Benchmarker has integration tests, which create actual cloud resources and take time and money to run. For this reason, they will only run when the variable PERFKIT_INTEGRATION is defined in the environment. The command

$ tox -e integration

will run the integration tests. The integration tests depend on having installed and configured all of the relevant cloud provider SDKs, and will fail if you have not done so.

Many... please add new requests via GitHub issues.