Testing Performance

This document verifies scheduling performance optimization in a Docker-based environment with bond4 and IPVLAN configurations.

The following uses the redis-benchmark stress test on four redis-server instances as an example to verify whether the feature is successfully enabled and to obtain the performance test results if the feature is enabled. This section includes affinity configuration and performance testing. Perform affinity configuration based on the container or physical machine scenario.

Affinity Configuration (Container Scenario)

Access the four Docker containers on the server.
1

docker exec -it {Container name} bash

Run the following command in the container:

        
             echo $nid > /sys/fs/cgroup/cpu/cpu.soft_domain

Domain-based scheduling binds cgroup to a NUMA node. nid identifies the NUMA node number using an offset of +1. For example, if the NUMA node number is 0, the value of nid is 1.

The CPU domain range defaults to the cluster level when domain-based scheduling is enabled. To configure the CPU domain range to the NUMA node level, set cpu.soft_domain_nr_cpu to the number of CPUs within the NUMA node.

          
               echo $numsCpu > /sys/fs/cgroup/cpu/cpu.soft_domain_nr_cpu

Start one redis-server instance in each Docker container. That is, four redis-server instances are started.

        
             cd path/redis-6.0.20 
./src/redis-server ./redis.conf --bind 0.0.0.0 --port 6379

Perform the test as instructed in Performance Testing.

Affinity Configuration (Physical Machine Scenario)

Run the following command on the physical machine:

        
             cd /sys/fs/cgroup/cpu
mkdir $dirname   # dirname is a user-defined directory.
echo $nid >/sys/fs/cgroup/cpu/$dirname/cpu.soft_domain

Domain-based scheduling binds cgroup to a NUMA node. nid identifies the NUMA node number using an offset of +1. For example, if the NUMA node number is 0, the value of nid is 1.

          
               echo $numsCpu > /sys/fs/cgroup/cpu/$dirname/cpu.soft_domain_nr_cpu

Start one redis-server instance in each Docker container. That is, four redis-server instances are started.

        
             cd path/redis-6.0.20
cgexec -g cpu:$dirname ./src/redis-server ./redis.conf --bind 0.0.0.0 --port 6379

Perform the test as instructed in Performance Testing.

Performance Testing

Run the following command to check whether domain-based scheduling is enabled:

        
             cat /sys/kernel/debug/sched_features | grep -e SOFT_DOMAIN

        
             cat /sys/fs/cgroup/cpu/docker/*/cpu.soft_domain

The command output is not empty, and the following information is displayed:

Go to the Redis directory on the client and prepare the stress test script. You can use the following script and modify parameters as needed.

Modify the following parameters as required:

REDIS_PATH: Redis root directory
REDIS_SERVER_IP_PREFIX: IP address subnet
redis_server_ip_suffix: start IP address suffix
instances: number of instances
client: parameter of the -c option, set to the optimal number of concurrent connections
size: parameter of the -d option, which is 3 (bytes) by default and can be changed to 256 or other values

#!/bin/bash 
 
REDIS_PATH="xxx1" # Redis directory
REDIS_PORT=6379 
REDIS_SERVER_IP_PREFIX="192.168.xx" 
redis_server_ip_suffix=128 # Start IP address suffix
instances=4 # Number of instances
client=200 # Parameter of -c
size=3  # Parameter of -d, which is 3 by default
 
# Stop the redis-benchmark process and clear test data logs.
pkill redis-benchmark 
DATA_LOG="xxx2" # Directory for storing performance test results
mkdir -p $DATA_LOG 
rm -rf ${DATA_LOG}/* 
 
# Perform the redis-benchmark stress test on the client.
job_ids="" 
for (( instance=1; instance<=instances; instance++ )); do 
    REDIS_SERVER_IP="${REDIS_SERVER_IP_PREFIX}.${redis_server_ip_suffix}" 
    echo "Running redis-benchmark on ${REDIS_SERVER_IP}:$REDIS_PORT" 
    echo "${REDIS_PATH}/src/redis-benchmark -h ${REDIS_SERVER_IP} -p $REDIS_PORT" 
    ${REDIS_PATH}/src/redis-benchmark -h ${REDIS_SERVER_IP} -p $REDIS_PORT -c $client -d $size -n 10000000 -r 10000000 -t set,get --threads 20 -q >> ${DATA_LOG}/${instances}_c${client}_d${size}_${REDIS_SERVER_IP}_${REDIS_PORT}.log & 
    job_ids="$job_ids $!" 
    ((redis_server_ip_suffix++)) 
done 
 
# Wait for redis-benchmark to complete.
echo "Waiting for the $instances jobs: SET, GET" 
wait $job_ids

Run the stress test script to perform the redis-benchmark test on the four instances simultaneously.
The performance test results are recorded in the DATA_LOG directory. Run the cat ./* command to view the results. The average performance of the four instances is the four-instance performance.