Skip to content

Gateway router with two CPUs

The table below displays the resource allocations for the Gateway related components of APK for the performance tests.

Container Name Requesting Memory Amount (Mi) Requesting CPU Amount (m) Limiting Memory Amount (Mi) Limiting CPU Amount (m)
Adapter 500 500 500 500
Enforcer 2000 1000 2000 1000
Router 1000 2000 (2 CPUs) 1000 2000 (2 CPUs)
Netty Backend 4096 2000 6114 2000

For the Enforcer's Java Virtual Machine (JVM), memory allocation was changed as below.

  • Xmx1500m (Maximum memory allocation for the JVM)
  • Xms1500m (Satrting memory allocation for the JVM)

In order to handle high request counts in Enforcer, authService's resource allocation changed as below.

[enforcer.authService]
    # Port of the Enforcer auth service
    port = 8081
    # Maximum message size in bytes
    maxMessageSize = 1000000000
    # Maximum header size in bytes
    maxHeaderLimit = 8192
    # Keep alive time in seconds for connection with the router via external authz service
    keepAliveTime = 600
    # Thread pool configurations of gRPC netty based server in Enforcer that handles the incoming requests in the Choreo Connect
    [enforcer.authService.threadPool]
        # Minimum number of workers to keep alive
        coreSize = 400
        # Maximum pool size
        maxSize = 1000
        # Timeout in seconds for idle threads waiting for work
        keepAliveTime = 600
        # Queue size of the worker threads
        queueSize = 2000

Throughput (requests/sec) vs. concurrent users

The following graph illustrates the throughput against the number of concurrent users.

TPS 2 CPUs

Key Observations

For a given payload size, APK provides a most likely consistent throughput level. But for large payload sizes you will get a low throughput value when compared to the small payload sizes.

Average Response time (ms) vs. concurrent users

Backend delay is assigned as zero milliseconds when obtaining test results. The below graph shows how response time varied for different concurrent user counts. Also, the same graph shows the impact of the payload size considering the same two parameters.

Response time 2 CPUs

Key observations:

When increasing concurrent users for given payload size, APK consumes steady growth for the response time. But when considering the same user count for large payloads, response time is high compared to the small payloads.

Response time percentiles vs. concurrent users

Below graphs show 90th, 95th, and 99th Response Time percentiles for 0ms backend delay. This is useful to measure the percentage of requests that exceeded the response time value for a given percentile. A percentile can also tell the percentage of requests completed below the particular response time value.

Percentiles 2 CPUs

Test scenario results in CSV format are available here.