Spring Webflux has huge performance drop from webFilter to controller

spring webflux version: 2.7.2

I tested a simple webflux project (although I tested with SSL/TLS, I believe the problem would still exist without them), and I found problems with titles like: "HTTP request returns immediately after entering webFilter" and "HTTP returns as soon as the request enters the controller", there is a huge performance difference between the two cases.

TPS tested with Apache Bench in the case of 4-cores CPU:

• "Return as soon as the HTTP request enters the webFilter" = 67000
• "The HTTP request will return immediately after entering the controller" = 20000 ~ 30000

Some additional conditions are added:

• In the webFilter, the only thing I do is add the keep-alive response header to the response, and then return the response. (code below)
• There is no code content that I added after the webFilter. The next piece of code I added is the controller
• no injection with @Autowired in the controller.
• The controller method uses @RequestBody to get parameters. The length of the input parameters is very small. Take the following code as an example: message 16 bytes (code below)

Code:

• MyFilter

public class MyFilter implements WebFilter {
  
  @Override
  public Mono<Void> filter(ServerWebExchange exchange, WebFilterChain chain) {
    exchange.getResponse().getHeaders().add("keep-alive", "true");
    
    // when i need to finish request
    return exchange.getResponse().writeWith(Mono.just(DefaultDataBufferFactory.sharedInstance.wrap("hello".getBytes())));

    // when i don't need to finish request
    return chain.filter(exchange);
  }
}

• MyController:

@RestController
public class MyController {
  
  @PostMapping("/hello")
  public Mono<HelloBody> hello(@RequestBody HelloBody helloBody) {
    return Mono.just(helloBody);
  }

  public static class HelloBody {
    private String message;

    // get/set
  }
}

As can be seen from the above code, such a test does not involve any business.

What I want to know is why there is such a big difference in performance returned by these two scenarios. More importantly, how can I reduce the performance loss of this part, so as to strive for more room for improvement in the performance of my program.