Skip to main content

Version: 3.2.16.4

Establish Performance Benchmark

To ensure the accuracy of API7 Gateway performance evaluation, please follow these key recommendations and steps before conducting benchmark tests.

Prerequisites

  • Select an Appropriate Number of Gateway Node:
    1. Use a single API7 Gateway node and configure the number of worker_processes based on machine resources (CPU cores or vCPU cores). It is recommended to initially configure 1 worker_processes for baseline testing and proceed with multi-core performance testing once the results are confirmed.
    2. Avoid using API7 Gateway nodes with multiple smaller configurations of worker_processes.
  • Gradually Increase the Number of worker_processes:
    1. During the initial testing, set the worker_processes to 1 to obtain a single-core performance baseline.
    2. After setting the single-core baseline performance, gradually increase the number of worker_processes to evaluate performance under multiple cores.
  • Exclude Upstream Network Interference:
    1. Enable only the mocking plugin to obtain the performance baseline of API7 Gateway. This plugin returns mock data in a specified format without forwarding requests to upstream servers.
  • Ensure Upstream Server Performance:
    1. Closely monitor the performance of API7 Gateway and upstream servers during testing to ensure that the upstream server is not a performance bottleneck.
  • Collect Baseline Results:
    1. Collect baseline results before testing additional scenarios. Configure 1 API7 Gateway node and enable 1 worker_processes;
    2. Eliminate gateway latency interference. Deploy the upstream server and API7 Gateway on the same machine and use Host networking;
    3. Ensure that the collected baseline results are largely consistent with the provided reference results.
  • Collect and Analyze Test Results:
    1. Collect multiple sets of test results and analyze the differences in the data using statistical methods (e.g., standard deviation) to ensure the stability and reliability of the test results.
  • Refer to Optimization Recommendations:
    1. After following the above recommendations and completing the tests, you can conduct benchmark tests for other scenarios based on actual needs. However, before doing so, please ensure that the collected performance baseline test results are largely consistent with the reference data, and carefully read the optimization suggestions provided below to make necessary adjustments to the configuration based on actual testing requirements.

Performance Baseline Test Deployment Topology

ec2 deploy

Performance Baseline Test Results

While collecting the baseline test results, API7 Gateway (with 1 worker_processes), the NGINX upstream server, and the benchmarking tool wrk2 were deployed on the same machine, utilizing the host network for communication.

Test ScenariosNumber of Routes/ConsumersQPSP99 (MS)P95 (MS)
No plugins enabled1 route, 0 consumers24,129.220.0930.082
No plugins enabled100 routes, 0 consumers23,652.910.0960.084
Only enable the limit-count1 route, 0 consumers20,495.100.1040.092
Only enable the limit-count100 routes, 0 consumers20,462.310.1040.094
Only enable the key-auth1 route, 1 consumer21,019.040.1000.089
Only enable the key-auth100 routes, 100 consumers20,444.810.1090.095
Enable both the key-auth and limit-count1 route, 1 consumer18,940.390.1100.097
Enable both the key-auth and limit-count100 routes, 100 consumers18,193.880.1100.098

By following the steps and recommendations outlined above, you will be able to more accurately evaluate the performance of API7 Gateway in different scenarios, providing valuable data support for subsequent optimization and deployment.

Optimize API7 Gateway Performance

Check Maximum Number of Open Files

Check the current system's maximum number of open files overall:

cat /proc/sys/fs/file-nr
3984 0 3255296

The last number 3255296 represents the maximum number of open files. If this number is too small on your machine, you need to modify the /etc/sysctl.conf file to increase it.

# /etc/sysctl.conf
fs.file-max = 1020000
net.ipv4.ip_conntrack_max = 1020000
net.ipv4.netfilter.ip_conntrack_max = 1020000

The change will take effect after a reload:

sudo sysctl -p /etc/sysctl.conf

Check the ulimit

Each user request corresponds to a file handle, and during stress testing, a large number of requests will be generated, need to increase this value. Use ulimit -n to check, and if it is a small value (the default is 1024), it should be modified to a million-level number, such as 1024000.

To temporarily modify it:

ulimit -n 1024000

To permanently modify it:

vim /etc/security/limits.conf

# Add the following lines
* hard nofile 1024000
* soft nofile 1024000

Avoiding Resource Contention

Ensure that wrk2, API7 Gateway, and upstream services are installed on separate machines and tested within the same host network to reduce the overhead caused by network latency.

If these resources are running on a single Kubernetes cluster, make sure the pods for the three services are scheduled on different nodes to avoid resource contention (e.g. in CPU and network), which can lead to degraded performance.

Upstream Server Reaching Limits

Check if the upstream server is reaching its limits. During the stress test, monitor the CPU and memory usage of the upstream server to determine if it has reached its maximum capacity. If increasing the worker_processes count in API7 Gateway does not yield any improvement, it may be that the upstream server or other systems have reached a bottleneck.

Disabling I/O for Access Logs

During performance benchmark testing, it is important to minimize the impact of access logs on performance, especially with high traffic scenarios involving significant log writing operations. Disabling access_log can alleviate the pressure on disk I/O.

Cloud Provider Performance Issues

Avoid using burstable instances provided by cloud servers. The available CPU for such instances is variable, which can unnecessarily impact test data.

info

The mapping between vCPUs (virtual CPU) and physical CPU provided by different cloud providers in their instance types is not always 1:1. Some instances use hyper-threading techniques, which means the actual number of physical CPU cores might be only half of the vCPUs count.

To accurately understand the vCPUs to physical CPU mapping for a specific instance type, it is recommended to consult the official documentation published by the cloud provider. For example, see AWS Instance Information.

API7 Gateway Internal Errors

Before starting performance benchmarking, it is important to adjust the log severity level in API7 Gateway to error and ensure that there are no internal errors in the error logs.

Using c1000k to Check Concurrent Connections

Install c1000k to check if the test environment can meet the requirements of 1 million concurrent connections.

# Start the server, listening on port 7000
. /server 7000

# Run the client to simulate a stress test
. /client 127.0.0.1 7000

API7 Gateway Configuration Example

config.yaml
nginx_config:
error_log_level: error
worker_processes: auto

http:
enable_access_log: false

API7.ai Logo

API Management for Modern Architectures with Edge, API Gateway, Kubernetes, and Service Mesh.

Product

API7 Cloud

SOC2 Type IIISO 27001HIPAAGDPRRed Herring

Copyright © APISEVEN Ltd. 2019 – 2024. Apache, Apache APISIX, APISIX, and associated open source project names are trademarks of the

Apache Software Foundation