You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@tomcat.apache.org by Christopher Schultz <ch...@christopherschultz.net> on 2025/09/29 16:57:11 UTC

Measuring response time

All,

There has been a few conversations and some nice work lately on adding 
nanosecond resolution to response-time measurements.

I wonder about what is being measured, though.

When a request is made, the TCP connection experiences several different 
phases:

1. Sitting in the OS's accept queue

2. Sitting in Tomcat's accept queue (this is after accept() returns and 
the request is added to our internal work-queue, but before a 
request-processing thread begins processing the request)

3. Being actively processed by a Tomcat service thread

I'm fairly sure that we are only actually measuring #3 when reporting 
response-time in access logs.

Would it be possible to add #2 as well?

I think it's impossible to add #1 because all of that happens before 
Java even gets a look at the request.

This may be more of a question of perception than anything else... while 
we /can/ measure response-time in ns, now... perhaps we are lying to 
ourselves about the actual experience of the client.

-chris


---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@tomcat.apache.org
For additional commands, e-mail: dev-help@tomcat.apache.org


Re: Measuring response time

Posted by Mark Thomas <ma...@apache.org>.
Hi Chris,

On 29/09/2025 17:57, Christopher Schultz wrote:
> All,
> 
> There has been a few conversations and some nice work lately on adding 
> nanosecond resolution to response-time measurements.
> 
> I wonder about what is being measured, though.
> 
> When a request is made, the TCP connection experiences several different 
> phases:
> 
> 1. Sitting in the OS's accept queue
> 
> 2. Sitting in Tomcat's accept queue (this is after accept() returns and 
> the request is added to our internal work-queue, but before a request- 
> processing thread begins processing the request)
> 
> 3. Being actively processed by a Tomcat service thread
> 
> I'm fairly sure that we are only actually measuring #3 when reporting 
> response-time in access logs.

Yes. Timing starts (for HTTP/1.1 anyway) when the first byte of the 
request is read.

> Would it be possible to add #2 as well?

Yes. Recording the request start time would need to move to the 
SocketWrapper but it could be done.

> I think it's impossible to add #1 because all of that happens before 
> Java even gets a look at the request.

Agreed.

> This may be more of a question of perception than anything else... while 
> we /can/ measure response-time in ns, now... perhaps we are lying to 
> ourselves about the actual experience of the client.

Actually doing something with the nanosecond resolution we now have is 
still on my TODO list. I question how useful it would be - even at 
microsecond resolution but the users asking for it are convinced it will 
be useful and as - so far - it has improved rather than reduced 
performance I'm not going to object.


Whether we want to include 2 in the request processing time is TBD. I'm 
not sure. I think a good first step would be to add some detail to the 
access log documentation so it is clear to folks exactly when timing 
starts and ends.

Mark


---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@tomcat.apache.org
For additional commands, e-mail: dev-help@tomcat.apache.org