diff options
Diffstat (limited to 'docs/testing/user/test-usage.rst')
| -rw-r--r-- | docs/testing/user/test-usage.rst | 422 |
1 files changed, 393 insertions, 29 deletions
diff --git a/docs/testing/user/test-usage.rst b/docs/testing/user/test-usage.rst index 8048cff..0fb3a6c 100644 --- a/docs/testing/user/test-usage.rst +++ b/docs/testing/user/test-usage.rst @@ -31,62 +31,137 @@ The typical test execution follows this pattern: #. Execute one or more performance runs #. Delete the environment -Configure The Environment -========================= +OpenStack or Stackless +====================== +StorPerf provides the option of controlling the OpenStack environment +via a Heat Stack, or it can run in stackless mode, where it connects +directly to the IP addresses supplied, regardless of how the slave +was created or even if it is an OpenStack VM. + +Note: Stack support in StorPerf will be deprecated as of the next release. + +Configure The Environment for OpenStack Usage +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The following pieces of information are required to prepare the environment: -- The number of VMs/Cinder volumes to create -- The Glance image that holds the VM operating system to use. StorPerf has - only been tested with Ubuntu 16.04 -- The OpenStack flavor to use when creating the VMs -- The name of the public network that agents will use -- The size, in gigabytes, of the Cinder volumes to create +- The number of VMs/Cinder volumes to create. +- The Cinder volume type (optional) to create +- The Glance image that holds the VM operating system to use. +- The OpenStack flavor to use when creating the VMs. +- The name of the public network that agents will use. +- The size, in gigabytes, of the Cinder volumes to create. +- The number of the Cinder volumes to attach to each VM. - The availability zone (optional) in which the VM is to be launched. Defaults to **nova**. - The username (optional) if we specify a custom image. - The password (optional) for the above image. +**Note**: on ARM based platforms there exists a bug in the kernel which can prevent +VMs from properly attaching Cinder volumes. There are two known workarounds: + +#. Create the environment with 0 Cinder volumes attached, and after the VMs + have finished booting, modify the stack to have 1 or more Cinder volumes. + See section on Changing Stack Parameters later in this guide. +#. Add the following image metadata to Glance. This will cause the Cinder + volume to be mounted as a SCSI device, and therefore your target will be + /dev/sdb, etc, instead of /dev/vdb. You will need to specify this in your + warm up and workload jobs. + +.. code-block: + --property hw_disk_bus=scsi --property hw_scsi_model=virtio-scsi + + The ReST API is a POST to http://StorPerf:5000/api/v1.0/configurations and takes a JSON payload as follows. .. code-block:: json - { - "agent_count": int, - "agent_flavor": string - "agent_image": string, - "public_network": string, - "volume_size": int, - "availability_zone": string, - "username": string, - "password": string - } + { + "agent_count": int, + "agent_flavor": "string", + "agent_image": "string", + "availability_zone": "string", + "password": "string", + "public_network": "string", + "username": "string", + "volume_count": int, + "volume_size": int, + "volume_type": "string" + } This call will block until the stack is created, at which point it will return -the OpenStack heat stack id. +the OpenStack heat stack id as well as the IP addresses of the slave agents. + + +Configure The Environment for Stackless Usage +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +To configure the environment for stackless usage, the slaves must be +fully operational (ie: a Linux operating system is running, are reachable +via TCP/IP address or hostname). + +It is not necessary to use the Configurations API, but instead define the +stack name as 'null' in any of the other APIs. This instructs StorPerf not to +gather information about the stack from OpenStack, and to simply use the +supplied IP addresses and credentials to communicate with the slaves. + +A slave can be a container (provided we can SSH to it), a VM running in any +hypervisor, or even a bare metal server. In the bare metal case, it even +allows for performing RADOS or RDB performance tests using the appropriate +FIO engine. -Initialize the Cinder Volumes +If the slave SSH server is listening to a port other than 22, the port number +can be specified as part of the address as follows: + +IPv4 example for port 2222: + +.. code-block:: + 192.168.1.10:2222 + +IPv6 example for port 2222: + +.. code-block:: + [1fe80::58bb:c8b:f2f2:c888]:2222 + +Helper Container Image for Workloads +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +A new docker container is provided with StorPerf that can be used to test +under docker or Kubernetes environments. It has hard coded credentials +of root/password with an SSH server built it, so be cautious about security +concerns when using this image. It listens internally on port 22, so that +port must be exposed to a free port on the host in order for StorPerf to +reach the synthetic workload container. + +.. code-block:: bash + + docker run --name=storperf-workloadagent -p 2222:22 + opnfv/storperf-workloadagent:latest + +Initialize the Target Volumes ============================= Before executing a test run for the purpose of measuring performance, it is -necessary to fill the Cinder volume with random data. Failure to execute this +necessary to fill the volume or file with random data. Failure to execute this step can result in meaningless numbers, especially for read performance. Most Cinder drivers are smart enough to know what blocks contain data, and which do not. Uninitialized blocks return "0" immediately without actually reading from the volume. -Initiating the data fill looks the same as a regular performance test, but uses -the special workload called "_warm_up". StorPerf will never push _warm_up -data to the OPNFV Test Results DB, nor will it terminate the run on steady state. -It is guaranteed to run to completion, which fills 100% of the volume with +Initiating the data fill behave similarly to a regular performance run, but +will tag the data with a special workload name called "_warm_up". It is +designed to run to completion, filling 100% of the specified target with random data. -The ReST API is a POST to http://StorPerf:5000/api/v1.0/jobs and -takes a JSON payload as follows. +The ReST API is a POST to http://StorPerf:5000/api/v1.0/initializations and +takes a JSON payload as follows. The body is optional unless your target +is something other than /dev/vdb. For example, if you want to profile a +glance ephemeral storage file, you could specify the target as "/filename.dat", +which is a file that then gets created on the root filesystem. .. code-block:: json { - "workload": "_warm_up" + "target": "/dev/vdb" } This will return a job ID as follows. @@ -100,6 +175,137 @@ This will return a job ID as follows. This job ID can be used to query the state to determine when it has completed. See the section on querying jobs for more information. +Authentication and Slave Selection +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +It is possible to run the Initialization API against a subset of the slaves +known to the stack, or to run it in stackless mode, where StorPerf +connects directly to the IP addresses supplied via SSH. The following +keys are available: + +slave_addresses + (optional) A list of IP addresses or hostnames to use as targets. If + omitted, and StorPerf is not running in stackless mode, the full list of + IP addresses from the OpenStack Heat stack is used. + +stack_name + (optional) Either the name of the stack in Heat to use, or null if running + in stackless mode. + +username + (optional) The username to supply to SSH when logging in. This defaults to + 'storperf' if not supplied. + +password + (optional) The password to supply to SSH when logging in. If omitted, the + SSH key is used instead. + +ssh_private_key + (optional) The SSH private key to supply to SSH when logging in. If omitted, + the default StorPerf private key is used. + +This shows an example of stackless mode going against a single bare metal +server reachable by IP address: + +.. code-block:: json + + { + "username": "labadmin", + "ssh_private_key": "-----BEGIN RSA PRIVATE KEY----- \nMIIE...X0=\n-----END RSA PRIVATE KEY-----", + "slave_addresses": [ + "172.17.108.44" + ], + "stack_name": null, + } + + +Filesystems and Mounts +~~~~~~~~~~~~~~~~~~~~~~ + +It is also possible to instruct StorPerf to create a file system on a device +and mount that as the target directory. The filesystem can be anything +supported by the target slave OS and it is possible to pass specific arguments +to the mkfs command. The following additional keys are available in the +Initializations API for file system control: + +mkfs + The type and arguments to pass for creating a filesystem + +mount_device + The target device on which to make the file system. The file system will + be mounted on the target specified. + +The following example shows the forced creation (-f) of an XFS filesystem +on device /dev/sdb, and mounting that device on /storperf/filesystem. + +**Note** If any of the commands (mkfs, mount) fail for any reason, the +Initializations API will return with a 400 code and the body of the response +will contain the error message. + +.. code-block:: json + + { + "target": "/storperf/filesystem", + "mkfs": "xfs -f", + "mount_device": "/dev/sdb", + } + + +Initializing Filesystems +~~~~~~~~~~~~~~~~~~~~~~~~ + +Just like we need to fill Cinder volumes with data, if we want to profile +files on a mounted file system, we need to initialize the file sets with +random data prior to starting a performance run. The Initializations API +can also be used to create test data sets. + +**Note** be sure to use the same parameters for the number of files, sizes +and jobs in both the Initializations API and the Jobs API, or you will end +up with possibly incorrect results in the Job performance run. + +The following keys are available in the Initializations API for file creation: + +filesize + The size of each file to be created and filled with random data. + +nrfiles + The number of files per job to create. + +numjobs + The number of independent instances of FIO to launch. + +Example: + +.. code-block:: json + + { + "target": "/storperf/filesystem", + "filesize": "2G", + "nrfiles": 10, + "numjobs": 10 + } + +This would create 100 (10 nrfiles x 10 numjobs) 2G files in the directory +/storperf/filesystem. + + +.. code-block:: json + + { + "username": "labadmin", + "ssh_private_key": "-----BEGIN RSA PRIVATE KEY----- \nMIIE...X0=\n-----END RSA PRIVATE KEY-----", + "slave_addresses": [ + "172.17.108.44" + ], + "stack_name": null, + "target": "/storperf/filesystem", + "mkfs": "ext4", + "mount_device": "/dev/sdb", + "filesize": "2G", + "nrfiles": 10, + "numjobs": 10 + } + + Execute a Performance Run ========================= Performance runs can execute either a single workload, or iterate over a matrix @@ -112,12 +318,152 @@ rr rs Read, Sequential. 100% read of sequential blocks of data rw - Read / Write Mix, Random. 70% random read, 30% random write + Read / Write Mix, Sequential. 70% random read, 30% random write wr Write, Random. 100% write of random blocks ws Write, Sequential. 100% write of sequential blocks. +Custom Workload Types +~~~~~~~~~~~~~~~~~~~~~ +New in Gambia (7.0), you can specify custom workload parameters for StorPerf +to pass on to FIO. This is available in the /api/v2.0/jobs API, and takes +a different format than the default v1.0 API. + +The format is as follows: + +.. code-block:: json + + "workloads": { + "name": { + "fio argument": "fio value" + } + } + +The name is used the same way the 'rr', 'rs', 'rw', etc is used, but can be +any arbitrary alphanumeric string. This is for you to identify the job later. +Following the name is a series of arguments to pass on to FIO. The most +important on of these is the actual I/O operation to perform. From the `FIO +manual`__, there are a number of different workloads: + +.. _FIO_IOP: http://git.kernel.dk/cgit/fio/tree/HOWTO#n985 +__ FIO_IOP_ + +* read +* write +* trim +* randread +* etc + +This is an example of how the original 'ws' workload looks in the new format: + +.. code-block:: json + + "workloads": { + "ws": { + "rw": "write" + } + } + +Using this format, it is now possible to initiate any combination of IO +workload type. For example, a mix of 60% reads and 40% writes scattered +randomly throughout the volume being profiled would be: + +.. code-block:: json + + "workloads": { + "6040randrw": { + "rw": "randrw", + "rwmixread": "60" + } + } + +Additional arguments can be added as needed. Here is an example of random +writes, with 25% duplicated blocks, followed by a second run of 75/25% mixed +reads and writes. This can be used to test the deduplication capabilities +of the underlying storage driver. + +.. code-block:: json + + "workloads": { + "dupwrite": { + "rw": "randwrite", + "dedupe_percentage": "25" + }, + "7525randrw": { + "rw": "randrw", + "rwmixread": "75", + "dedupe_percentage": "25" + } + } + +There is no limit on the number of workloads and additional FIO arguments +that can be specified. + +Note that as in v1.0, the list of workloads will be iterated over with the +block sizes and queue depths specified. + +StorPerf will also do a verification of the arguments given prior to returning +a Job ID from the ReST call. If an argument fails validation, the error +will be returned in the payload of the response. + +File System Profiling +~~~~~~~~~~~~~~~~~~~~~ + +As noted in the Initializations API, files in a file system should be +initialized prior to executing a performance run, and the number of jobs, +files and size of files should match the initialization. Given the following +Initializations API call: + +.. code-block:: json + + { + "username": "labadmin", + "ssh_private_key": "-----BEGIN RSA PRIVATE KEY----- \nMIIE...X0=\n-----END RSA PRIVATE KEY-----", + "slave_addresses": [ + "172.17.108.44" + ], + "stack_name": null, + "target": "/storperf/filesystem", + "mkfs": "ext4", + "mount_device": "/dev/sdb", + "filesize": "2G", + "nrfiles": 10, + "numjobs": 10 + } + +The corresponding call to the Jobs API would appear as follows: + +.. code-block:: json + + { + "username": "labadmin", + "ssh_private_key": "-----BEGIN RSA PRIVATE KEY----- \nMIIE...X0=\n-----END RSA PRIVATE KEY-----", + "slave_addresses": [ + "172.17.108.44" + ], + "stack_name": null, + "target": "/storperf/filesystem", + "block_sizes": "4k", + "queue_depths": "8", + "workloads": { + "readwritemix": { + "rw": "rw", + "filesize": "2G", + "nrfiles": "10", + "numjobs": "10" + } + } + } + +**Note** the queue depths and block sizes as well as the I/O pattern (rw) +can change, but the filesize, nrfiles, numjobs and slave addresses must +match the initialization or the performance run could contain skewed results +due to disk initialization. StorPerf explicitly allows for the mismatch +of these so that it is possible to visualize performance when the files +or disks have not been properly initialized. + + Block Sizes ~~~~~~~~~~~ A comma delimited list of the different block sizes to use when reading and @@ -165,6 +511,24 @@ is required in order to push results to the OPNFV Test Results DB: "test_case": "snia_steady_state" } +Changing Stack Parameters +~~~~~~~~~~~~~~~~~~~~~~~~~ +While StorPerf currently does not support changing the parameters of the +stack directly, it is possible to change the stack using the OpenStack client +library. The following parameters can be changed: + +- agent_count: to increase or decrease the number of VMs. +- volume_count: to change the number of Cinder volumes per VM. +- volume_size: to increase the size of each volume. Note: Cinder cannot shrink volumes. + +Increasing the number of agents or volumes, or increasing the size of the volumes +will require you to kick off a new _warm_up job to initialize the newly +allocated volumes. + +The following is an example of how to change the stack using the heat client: + +.. code-block:: + heat stack-update StorPerfAgentGroup --existing -P "volume_count=2" Query Jobs Information |