HttpServer performance issue / improvement

[Daniel Jeliński]

Hi Robert,
If you are on Linux, see test/jdk/java/net/vthread/HttpALot.java. That
test was recently modified to add -Dsun.net.httpserver.nodelay=true
[1], because it took forever to complete on LInux machines. On other
OSes the problem is not visible in existing tests.
Cheers,
Daniel

[1] https://github.com/openjdk/jdk/pull/10504

sob., 6 kwi 2024 o 00:37 robert engels <rengels at ix.netcom.com> napisał(a):
>
> Hi Daniel,
>
> I think I have a solution that would work. I will try to get a PR together. Do you know if there is an existing test case the demonstrates the issue? - if not, I will start with that.
>
> Robert
>
> On Apr 4, 2024, at 9:44 AM, Daniel Jeliński <djelinski1 at gmail.com> wrote:
>
> Hi Robert,
> Thanks for bringing this up! We are aware of the issue, it's tracked under https://bugs.openjdk.org/browse/JDK-6968351.
>
> If you have an idea for a proper fix that doesn't add too much complexity, please open a PR, and we'll be happy to help.
> Cheers,
> Daniel
>
> czw., 4 kwi 2024, 14:55 użytkownik Robert Engels <rengels at ix.netcom.com> napisał:
>>
>> 
>> When doing some testing on github.com/robaho/httpserver - which is a fork of the jdk http server, I discovered a significant performance issue.
>>
>> When an http connection is in ‘keep-alive’ - the default for http 1.1 - the headers are “flushed” here https://github.com/openjdk/jdk21/blob/890adb6410dab4606a4f26a942aed02fb2f55387/src/jdk.httpserver/share/classes/sun/net/httpserver/ExchangeImpl.java#L281
>>
>> This means that after the handler runs and it sends data - e.g. /hello sends “hello” on the connection, the connection will stall due to the Nagel algorithm - usually incurring a 50 ms delay. The stall occurs since the client will not see the expected data until after the delay, so it is unable to send the next (when reusing the same connection/HttpClient).
>>
>> You can set the TCP_NODELAY on the server to work-around this, but a better solution would be to override the flush() on the BufferedOutputStream to not flush() the underlying connection - i.e. only write the buffered bytes, or rework it a bit to only flush when there is no content to send.
>>
>