neo4j-streams performance test

This part of the project aims to give a standard way to understand the performances of the platform and highlight the configuration results.

Currently, it's available only the test of neo4j to neo4j configuration, in remote too.

Setup

The basic unit of work used to test the performances needs a specific cypher part in order to get the timestamp of creation and receiving of the event. The elapsed time is calculated using these timestamps through cypher queries using: Elapsed = max(tc) - min(tp) where tc is the instant of insert in the cunsumer and tp the instant of insert in the producer.

Kafka

Default kafka configuration, running on sigle instance on the same machine of the other components

• Kafka 2.11-0.10.1.1

Producer

• neo4j 3.4.7 community edition
• APOC 3.4.0.3
• 8 GB heap memory
• CREATE INDEX ON :Performance(group)
• default neo4j-streams configuration (without filtering)
• kafka.batch.size=16384

The creation query used by the tester:

UNWIND range(1,{msg}) as ran
CREATE (n:Performance)
SET n.group = {uuid}, n.creation_time = apoc.date.currentTimestamp()
RETURN min(n.creation_time) as creation_time

Consumer

• neo4j 3.4.7 community edition
• APOC 3.4.0.3
• 2 GB heap memory
• CREATE INDEX ON :Performance(group)
• default neo4j-streams configuration
• kafka.max.poll.records=16384

The acquisition query:

streams.sink.topic.cypher.neo4j=WITH event.payload AS payload, event.meta AS meta CALL apoc.do.case( [\

payload.type = 'node' AND meta.operation = 'created', \

'CREATE (x:Performance {received_time: apoc.date.currentTimestamp()}) SET x+=props RETURN count(x)'] \

,'RETURN 0', \

{props:  payload.after.properties}) \

YIELD value RETURN count(value)

Run the tests

The tester is written in Python. Check your system for:

• Python 3.6
• py2neo
• matplotlib

config.ini

To run the test you need to edit the config.ini file in order to setup the connection data for the producer and consumer neo4j instances.

The unit part is used as default values to run the tests.

• repeat is the number of times each test is executed
• node is the number of nodes created for each test

[producer]
url = bolt://localhost:8687
username = neo4j
password = producer

[consumer]
url = bolt://localhost:7687
username = neo4j
password = consumer

[unit]
repeat = 5
nodes = 1000

start

To check the coniguration or to get a fast result you can run the command

python3 neo4j-streams-pt.py --start

The output is a windows that show you the distribution of the test. See result session to better understand the values. You can use the option --plot-out file.png to not show the result but save it on file. The details of the execution are dumped on standard output as CSV or you can redirect it on file using the option --csv-out file.csv

baseline

This is the main command you have to test the system. By command line arguments you can specify the series to test. The arguments are the number of node's unit to use.

ES: 3 = 3 * 1000 if 1000 is the value configured as nodes in the [unit] part.

To get the same result of start:

python3 neo4j-streams-pt.py --baseline 1

The get more distribution (real case):

python3 neo4j-streams-pt.py --baseline 1 10 100 1000 --plot-out results.png

This means the tester execute 5 (if repeat = 5) times each series. A series is comped by 1k, 10k, 100k, 1000k nodes (if nodes = 1000)

Results

There're lot of variables to tuning the whole process and lot of scenarios to optimize (such as high or low volumes of events per transaction). We introduce here two cases, both on a local machine, one with docker and one without. The computer is a MacBook Pro with

• CPU 2,2 GHz Intel Core i7
• 16 GB DDR3, 1600 MHz
• SSD

With Docker

Here's the results with the setup as described above with

• Docker Community Edition 18.06.1-ce-mac73
• docker-compose.yml

As you can see, there's not a single number to describe the scenario, but 0.20 ms per node (for the simple CREATE query) colud be a reference time.

Without Docker

Here's the results with the setup as described above with

• neo4j producer, neo4j consumer, Kafka and Zookeeper run on the same machine, the tester too. The clock is the same, so it's synchronized.
• Kafka and Zookeeper run with default configuration

As you can see, there's not a single number to describe the scenario, but 0.13 ms per node (for the simple CREATE query) colud be a reference time.