MultiNode API

The LAVA MultiNode API provides a simple way to pass messages between devices in a group, using the connection which is already available through LAVA. The API is not intended for transfers of large amounts of data. Test definitions which need to transfer files, long messages or other large amounts of data need to set up their own network configuration, access and download methods and do the transfer in the test definition.

Guidance in using the API

It is recommended to avoid doing a lot of calculation within the calls to the API. There are times when a script is needed to retrieve data from the test shell, but avoid running that script in the call to the API. Always check the output of the script (e.g. with lava-test-case) and/or run the script separately in the test definition run steps so that the output appears in the test job logs. Preparing and outputting the data before sending it with the API will aid in debugging the test definition.

Note

Debugging of complex test definitions does not only happen during the initial development. There may be further issues and corner cases, uncovered only after a test job was in use for a while. Retain enough structure in your test definitions to be able to debug problems later without needing to resubmit the MultiNode test job (as some problems may be non-deterministic, etc.).

Note

It is not recommended to use lava-test-case command in conjunction with the MultiNode API calls. The first reason is that any errors that might occur within the API will be ignored by the lava-test-case and it will be seen as successful by lava-test-shell. The second reason is that the job will end up with duplicate test cases for each API call (one from lava-test-case and the other one from API command).

lava-self

Usage:

lava-self

lava-self reports the job ID, as the dispatcher itself has no knowledge of the hostname of the deployed system or the original database name of the device. The output of lava-group can also be used. (This behavior changed in the 2017.9 release.)

lava-role

Usage:

lava-role

Prints the role the current device is playing in a MultiNode job.

Example. In a directory with several scripts, one for each role involved in the test:

$ ./run-$(lava-role)

Usage:

lava-role list

Prints a list of all roles within this MultiNode job, separated by whitespace.:

#!/bin/sh
for role in `lava-role list`; do
    echo $role
done

lava-group

Usage:

lava-group

This command will produce in its standard output a representation of the device group that is participating in the MultiNode test job.

The output format contains one line per device, and each line contains the job ID and the role that job is playing in the test, separated by a TAB character:

12345     client
12346     loadbalancer
12347     backend
12348     backend

Caution

This behavior changed in 2017.9 as V2 does not have knowledge of the device hostname, only the job ID for each role.

Usage:

lava-group <role>

This command will produce in its standard output a list of the test jobs assigned the specified role in the MultiNode test job.

The output format contains one line per job ID assigned to the specified role. The name of the role itself is not printed:

$ lava-group client
12345
$ lava-group backend
12347
12348

If there is no matching role, exits with non-zero status code and outputs nothing:

$ lava-group server ; echo $?
1

If your test definition relies on a particular role, one of the first test cases should be to check this role has been defined:

- lava-test-case check-server-role --shell lava-group server

lava-send

Sends a message to the group, optionally passing associated key-value data pairs. Sending a message is a non-blocking operation. The message is guaranteed to be available to all members of the group, but some of them might never retrieve it.

The message-id will be persistent for the lifetime of the target group managing the entire multinode test job. Re-sending a different message with an existing message-id is not supported.

Usage:

lava-send <message-id> [key1=val1 [key2=val2] ...]

Examples are provided below, together with lava-wait and lava-wait-all.

lava-wait

Waits until any device in the group sends a message with the given ID. This call will block until such message is sent.

Usage:

lava-wait <message-id>

If there was data passed in the message, the key-value pairs will be stored in the cache file (/tmp/lava_multi_node_cache.txt by default), each in one line. If no key-values were passed, nothing is stored.

The message ID data is persistent for the life of the MultiNode group. The data can be retrieved at any later stage using lava-wait and as the data is already available, there will be no waiting time for repeat calls. If devices continue to send data with the associated message ID, that new data will continue to be added to the stored data for that message ID and will be returned by subsequent calls to lava-wait for that message ID. Use different message ID(s) if you don’t want this effect.

lava-wait-all

lava-wait-all operates in different ways, depending on the presence of the role parameter.

lava-wait-all <message-id> [<role>]

If data was sent by the devices with the message, the key-value pairs will be stored in the cache file (/tmp/lava_multi_node_cache.txt by default), each in one line, prefixed with the target name and a colon.

Some examples for lava-send, lava-wait and lava-wait-all are given below.

The message returned can include data from devices which sent a message with the relevant message ID, only the wait is dependent on particular devices with a specified role.

As with lava-wait, the message ID is persistent for the duration of the MultiNode group.

lava-wait-all <message-id>

lava-wait-all <message-id>

lava-wait-all waits until all devices in the group send a message with the given message ID. Every device in the group must use lava-send with the same message ID for lava-wait-all to finish, or any device using this API call will wait forever (and eventually timeout, failing the job).

Using lava-sync or lava-wait-all in a test definition effectively makes all boards in the group run at the speed of the slowest board in the group up to the point where the sync or wait is called.

lava-wait-all <message-id> <role>

lava-wait-all <message-id> <role>

If <role> is used, only wait until all devices with that given role send a message with the matching message ID. Devices of the given role do not enter lava-wait, but just send the message and continue the test definition. Ensure the test continues for long enough for the devices using lava-wait-all to pick up the message and act on it. Typically, this involves using a lava-sync after the lava-send on devices with the given role and after the completion of the task on the devices which were waiting for the message.

Not all roles in the group need to send a message or wait for a message. One role will act as a sender, at least one role will act as a receiver and any other roles can continue as normal. Note that this level of fine-grained control is usually not needed. It is advisable to draw out the sequence in a table to ensure that the correct calls are made.

lava-sync

Global synchronization primitive. Sends a message, and waits for the same message from all of the other devices.

Usage:

lava-sync <message>

lava-sync foo is effectively the same as lava-send foo followed by lava-wait-all foo.

A lava test result is generated within the current Test Suite, recording the completion or failure of the synchronization.

Example 1: Simple client-server MultiNode test

Two devices, with roles client, server

LAVA Test Shell test definition (say, example1.yaml):

run:
    steps:
        - ./run-`lava-role`.sh

The test image or the test definition would then provide two scripts, with only one being run on each device, according to the role specified.

run-server.sh:

#!/bin/sh

SPACE=`df -h | grep "/$" | awk '{print $4}'`
echo $SPACE
lava-send server-ready free-space=$SPACE

Notes:

  • To make use of the server-ready message, some kind of client needs to do a lava-wait server-ready

run-client.sh:

#!/bin/sh

lava-wait server-ready
free-space=$(cat /tmp/lava_multi_node_cache.txt | cut -d = -f 2)
echo "The free disk space on server is ${free-space}"

Notes:

  • The client waits for the server-ready message then get the data which was sent by server from /tmp/lava_multi_node_cache.txt

Example 2: iperf client-server test

Two devices, with roles client, server

LAVA Test Shell test definition (say, example1.yaml):

run:
    steps:
        - ./run-`lava-role`.sh

The test image or the test definition would then provide two scripts, with only one being run on each device, according to the role specified.

run-server.sh:

#!/bin/sh

iperf -s &
echo $! > /tmp/iperf-server.pid
IP=`ip route get 8.8.8.8 | head -n 1 | awk '{print $NF}'`
echo $IP
lava-send server-ready server-ip=$IP
lava-wait client-done
kill -9 `cat /tmp/iperf-server.pid`

Notes:

  • iperf server process needs to be run in the background to wait for the connection from the client and the process id will be stored somewhere for later use.

  • To make use of the server-ready message, some kind of client needs to do a lava-wait server-ready

  • There needs to be a support on a client to do the lava-send client-done or the server role will fail with a timeout.

  • If there was more than one client, the server could call lava-wait-all client-done instead.

  • iperf server process must be killed after getting client-done message, otherwise the test job will not proceed.

run-client.sh:

#!/bin/sh

lava-wait server-ready
server=$(cat /tmp/lava_multi_node_cache.txt | cut -d = -f 2)
iperf -c $server
# ... do something with output ...
lava-send client-done

Notes:

  • The client waits for the server-ready message as its first task, then does some work, then sends a “done” message so that the server can move on and do other tests.

Example 3: variable number of clients

run-server.sh:

#!/bin/sh

start-server
lava-sync ready
lava-sync done

run-client.sh:

#!/bin/sh

# refer to the server by name, assume internal DNS works
server=$(lava-group | grep 'server$' | cut -f 1)

lava-sync ready
run-client
lava-sync done

Example 4: peer-to-peer application

Single role: peer, any number of devices

run-peer.sh:

#!bin/sh

initialize-data
start-p2p-service
lava-sync running

push-data
for peer in $(lava-group | cut -f 1); do
    if [ $peer != $(lava-self) ]; then
        query-data $peer
    fi
done

Using a flow table to plan the job

Synchronization of any type needs to be planned and the simplest way to manage the messages between roles within a group is to set out a strict table of the flow.

Set out the call and leave blank rows until that call is matched by the appropriate roles, to represent the time that the devices with that role will block in a wait loop with the coordinator.

Server

Client

Observer

deploy & boot

deploy & boot

deploy & boot

lava-sync start

lava-sync start

lava-sync start

server_start.sh

lava-wait-all ready server

lava-sync fin

lava-send ready

lava-sync fin

client-tasks.sh

lava-sync fin

In this overly simplistic table, the Observer role really has nothing useful to do but to demonstrate that it will spend most of it’s time in lava-sync fin.

All roles will wait in lava-sync start until all deploy and boot operations (or whatever other tasks are put ahead of the call to lava-sync) are complete. The flow table does not include this delay.

The Server role runs a script to start a service, sending “ready” when the script returns.

The Client role waits until all devices with the Server role have completed lava-send ready. Observer is unaffected and Server moves directly into the lava-sync fin. Once the Client completes lava-wait-all ready server, the Client can run the client tasks script. That script finally puts the devices with the Client role into lava-sync fin at which point, the Client role receives the message that everyone else is already in that sync, the sync completes and the flow table ends.

Tables like this also help visualize how long the timeouts need to be to allow the Observer role to wait for all the server tasks and all the client tasks to complete.