Benchmark an Auto Tiering enabled database

Auto Tiering on Redis Enterprise Software lets you use cost-effective Flash memory as a RAM extension for your database.

But what does the performance look like as compared to a memory-only database, one stored solely in RAM?

These scenarios use the memtier_benchmark utility to evaluate the performance of a Redis Enterprise Software deployment, including the trial version.

The memtier_benchmark utility is located in /opt/redislabs/bin/ of Redis Enterprise Software deployments. To test performance for cloud provider deployments, see the memtier-benchmark GitHub project.

For additional, such as assistance with larger clusters, contact support.

Benchmark and performance test considerations

These tests assume you're using a trial version of Redis Enterprise Software and want to test the performance of a Auto Tiering enabled database in the following scenarios:

• Without replication: Four (4) master shards
• With replication: Two (2) primary and two replica shards

With the trial version of Redis Enterprise Software you can create a cluster of up to four shards using a combination of database configurations, including:

• Four databases, each with a single master shard
• Two highly available databases with replication enabled (each database has one master shard and one replica shard)
• One non-replicated clustered database with four master shards
• One highly available and clustered database with two master shards and two replica shards

Test environment and cluster setup

For the test environment, you need to:

1. Create a cluster with three nodes.
2. Prepare the flash memory.
3. Configure the load generation tool.

Creating a three-node cluster

This performance test requires a three-node cluster.

You can run all of these tests on Amazon AWS with these hosts:

• 2 x i3.2xlarge (8 vCPU, 61 GiB RAM, up to 10GBit, 1.9TB NMVe SSD)

  These nodes serve RoF data

• 1 x m4.large, which acts as a quorum node

To learn how to install Redis Enterprise Software and set up a cluster, see:

• Redis Enterprise Software quickstart for a test installation
• Install and upgrade for a production installation

These tests use a quorum node to reduce AWS EC2 instance use while maintaining the three nodes required to support a quorum node in case of node failure. Quorum nodes can be on less powerful instances because they do not have shards or support traffic.

As of this writing, i3.2xlarge instances are required because they support NVMe SSDs, which are required to support RoF. Auto Tiering requires Flash-enabled storage, such as NVMe SSDs.

For best results, compare performance of a Flash-enabled deployment to the performance in a RAM-only environment, such as a strictly on-premises deployment.

Prepare the flash memory

After you install RS on the nodes, the flash memory attached to the i3.2xlarge instances must be prepared and formatted with the /opt/redislabs/sbin/prepare_flash.sh script.

Set up the load generation tool

The memtier_benchmark load generator tool generates the load on the RoF databases. To use this tool, install RS on a dedicated instance that is not part of the RS cluster but is in the same region/zone/subnet of your cluster. We recommend that you use a relatively powerful instance to avoid bottlenecks at the load generation tool itself.

For these tests, the load generation host uses a c4.8xlarge instance type.

Database configuration parameters

Create a Auto Tiering test database

You can use the Redis Enterprise Cluster Manager UI to create a test database. We recommend that you use a separate database for each test case with these requirements:

Data population

Populate the benchmark dataset

The memtier_benchmark load generation tool populates the database. To populate the database with N items of 500 Bytes each in size, on the load generation instance run:

$ memtier_benchmark -s $DB_HOST -p $DB_PORT --hide-histogram
--key-maximum=$N -n allkeys -d 500 --key-pattern=P:P --ratio=1:0

Set up a test database:

Centralize the keyspace

With replication

To create roughly 20.5 million items in RAM for your highly available clustered database with 75 million items, run:

$ memtier_benchmark  -s $DB_HOST -p $DB_PORT --hide-histogram
--key-minimum=27250000 --key-maximum=47750000 -n allkeys
--key-pattern=P:P --ratio=0:1

To verify the database values, use Values in RAM metric, which is available from the Metrics tab of your database in the Cluster Manager UI.

Without replication

To create 41 million items in RAM without replication enabled and 150 million items, run:

$ memtier_benchmark  -s $DB_HOST -p $DB_PORT --hide-histogram
--key-minimum=54500000 --key-maximum=95500000 -n allkeys
--key-pattern=P:P --ratio=0:1

Test runs

Generate load

With replication

We recommend that you do a dry run and double check the RAM Hit Ratio on the Metrics screen in the Cluster Manager UI before you write down the test results.

To test RoF with an 85% RAM Hit Ratio, run:

$ memtier_benchmark -s $DB_HOST -p $DB_PORT --pipeline=11 -c 20 -t 1
-d 500 --key-maximum=75000000 --key-pattern=G:G --key-stddev=5125000
--ratio=1:1 --distinct-client-seed --randomize --test-time=600
--run-count=1 --out-file=test.out

Without replication

Here is the command for 150 million items:

$ memtier_benchmark -s $DB_HOST -p $DB_PORT --pipeline=24 -c 20 -t 1
-d 500 --key-maximum=150000000 --key-pattern=G:G --key-stddev=10250000
--ratio=1:1 --distinct-client-seed --randomize --test-time=600
--run-count=1 --out-file=test.out

Where:

Test results

Monitor the test results

You can either monitor the results in the Metrics tab of the Cluster Manager UI or with the memtier_benchmark output. However, be aware that:

• The memtier_benchmark results include the network latency between the load generator instance and the cluster instances.

• The metrics shown in the Cluster Manager UI do not include network latency.

Expected results

You should expect to see an average throughput of:

• Around 160,000 ops/sec when testing without replication (i.e. Four master shards)
• Around 115,000 ops/sec when testing with enabled replication (i.e. 2 master and 2 replica shards)

In both cases, the average latency should be below one millisecond.