Memory surge during load testing

[jinzaz]

Describe the bug

When I created an empty Web API Project and only added the reverse proxy service, and then used JMeter for load testing, I found a surge in memory. Below was the stack information monitored using dotnet dump and dotnet counters. I suspected that Jeeter was using keep-alive, but I added the app UseRequestTimeouts(); There will still be these phenomena, and I don't know what causes them

Further technical details

• yarp.reverse-proxy version 2.2.0-preview.1.24266.1
• linux docker contianer
• net 8.0

[MihaZupan]

Was the memory dump screenshot you shared taken after all the requests already completed?

Also just in case, can you share how you're using YARP (e.g. config file) / whether and how you're using the IHttpForwarder directly.

[jinzaz]

yeah this dump was after all the requests already completed!，this is my YARP config

  "ReverseProxy": {
    "Routes": {
      "basic-apiservice": {
        "Match": {
          "Methods": null,
          "Hosts": null,
          "Path": "/xxx/xxxx/{**catch-all}",
          "QueryParameters": null,
          "Headers": null
        },
        "ClusterId": "basic-apiservice",
        "AuthorizationPolicy": null,
        "RateLimiterPolicy": null,
        "Timeout": "00:00:20",
        "CorsPolicy": null,
        "Metadata": null,
        "Transforms": [
          {
            "PathRemovePrefix": "/xxxx/xxxx"
          }
        ]
      }
    },
    "Clusters": {
      "basic-apiservice": {
        "LoadBalancingPolicy": null,
        "SessionAffinity": null,
        "HealthCheck": null,
        "HttpClient": {
          "EnableMultipleHttp2Connections": true
        },
        "HttpRequest": {
          "AllowResponseBuffering": "false"
        },
        "Destinations": {
          "basic-apiservice/destination1": {
            "Address": "http://xxxx",
            "Health": "http://xxxxxx",
            "Metadata": null,
            "Host": null
          }
        },
        "Metadata": null
      }
    }
  }

I seem to have found a solution， I seem to have found the cause of the problem, partly due to the PinnedBlockMemoryPool in aspnetcore when the container memory is not restricted. This is similar to this problem ，dotnet/aspnetcore#24958
and partly due to the low execution frequency of GC Server Mode in GC mode，But now I'm solving it by using GarbageCollectionAdaptationMode

[jinzaz]

But I found that YARP's CPU usage is very high, and I'm not sure if this is normal, or can you recommend optimization methods

[adityamandaleeka]

Can you specify more about the CPU usage being high? Is it high under load relative to when idle? That is expected. In general, CPU traces are a useful way to determine what's happening with a process when it is using a significant amount of CPU time.

[jinzaz]

Under actual load, the CPU usage is relatively high because the upstream of my load also passes through nginx, but nginx's CPU usage is much lower. Therefore, I want to know if this is normal

[doddgu]

Under actual load, the CPU usage is relatively high because the upstream of my load also passes through nginx, but nginx's CPU usage is much lower. Therefore, I want to know if this is normal

I also found a problem with high cpu usage, which can be seen #2427
This is a very serious problem in heavy traffic.

[MihaZupan]

Has using the GC adaptation mode resolved your memory usage concerns?
The feature is built for scenarios like this, where you want to reclaim memory that was only needed during a temporary surge in traffic. Note also that it will be enabled by default as of .NET 9 with the server GC.

Re: CPU consumption, are your request latencies impacted by it? We're aware of scenarios where CPU usage may stay high as we're expecting to deal with more traffic. See e.g. dotnet/runtime#72153 (comment)
Do you see a positive change if you set DOTNET_ThreadPool_UnfairSemaphoreSpinLimit=0?

[doddgu]

Re: CPU consumption, are your request latencies impacted by it? We're aware of scenarios where CPU usage may stay high as we're expecting to deal with more traffic. See e.g. dotnet/runtime#72153 (comment) Do you see a positive change if you set DOTNET_ThreadPool_UnfairSemaphoreSpinLimit=0?

@MihaZupan hi, thanks for your reply.

Based on our current observations, the increase in backend service latency affects the number of threads in the thread pool and the CPU usage. The number of requests being processed is also significantly increasing, and the final latency will inevitably be impacted. However, from the surface analysis, I think this might be due to the temporary creation of a large number of threads. Of course, we will try your suggestion and will continue to provide feedback once we have results.

[zhenlei520]

Has using the GC adaptation mode resolved your memory usage concerns? The feature is built for scenarios like this, where you want to reclaim memory that was only needed during a temporary surge in traffic. Note also that it will be enabled by default as of .NET 9 with the server GC.

Re: CPU consumption, are your request latencies impacted by it? We're aware of scenarios where CPU usage may stay high as we're expecting to deal with more traffic. See e.g. dotnet/runtime#72153 (comment) Do you see a positive change if you set DOTNET_ThreadPool_UnfairSemaphoreSpinLimit=0?

We tried to optimize CPU usage by configuring the following environment variables
DOTNET_ThreadPool_ThreadsToKeepAlive=500,DOTNET_ThreadPool_UnfairSemaphoreSpinLimit=0,DOTNET_SYSTEM_NET_SOCKETS_INLINE_COMPLETIONS=1
After the modification, the thread accumulation situation improved. Modifying the DOTNET_ThreadPool_UnfairSemaphoreSpinLimit value helped improve CPU utilization. Compared with before, the CPU was reduced by 30%. However, there was still a sudden high CPU usage. However, at this time, the downstream response time did not fluctuate significantly, and the total number of requests did not change significantly, but the processing capacity slowed down. The good news is that it quickly dropped. It will take some time to observe the specific reasons.

[zhenlei520]

Has using the GC adaptation mode resolved your memory usage concerns? The feature is built for scenarios like this, where you want to reclaim memory that was only needed during a temporary surge in traffic. Note also that it will be enabled by default as of .NET 9 with the server GC.
Re: CPU consumption, are your request latencies impacted by it? We're aware of scenarios where CPU usage may stay high as we're expecting to deal with more traffic. See e.g. dotnet/runtime#72153 (comment) Do you see a positive change if you set DOTNET_ThreadPool_UnfairSemaphoreSpinLimit=0?

We tried to optimize CPU usage by configuring the following environment variables DOTNET_ThreadPool_ThreadsToKeepAlive=500,DOTNET_ThreadPool_UnfairSemaphoreSpinLimit=0,DOTNET_SYSTEM_NET_SOCKETS_INLINE_COMPLETIONS=1 After the modification, the thread accumulation situation improved. Modifying the DOTNET_ThreadPool_UnfairSemaphoreSpinLimit value helped improve CPU utilization. Compared with before, the CPU was reduced by 30%. However, there was still a sudden high CPU usage. However, at this time, the downstream response time did not fluctuate significantly, and the total number of requests did not change significantly, but the processing capacity slowed down. The good news is that it quickly dropped. It will take some time to observe the specific reasons.

After testing, it was found that modifying DOTNET_ThreadPool_ThreadsToKeepAlive would cause greater CPU fluctuations, and it is better to let the system manage the thread pool itself than to control it

[MihaZupan]

The memory usage aspect appears to be addressed by using the GC adaptation mode (which is enabled by default as of .NET 9).

I'll close this issue in favor of continuing the CPU usage discussion in #2427.