Re: [libmicrohttpd] RE: MHD_handle_connection() performance issue

[Christian Grothoff]

On Wednesday, March 23, 2011 02:06:23 am Eivind Sarto wrote:
> I going to give some feedback to my own posting :-)
> 
> I think the solution is that MHD_handle_connection() should not set the
> POLLIN flag for the listen/server socket. It should change the code to the
> following:
>            fd = con->daemon->socket_fd;
>            p[1].fd = fd;
> //         p[1].events |= POLLIN;     /*OLD*/
>            p[1].events = 0;                  /*NEW*/
> 
> Even if pollfd.events=0, pollfd.revents=POLLHUP when the daemon socket is
> shout down.   And that is what we care about, right? We don't want to know
> if there is any input data on the daemon socket.

That is certainly right.
 
> I looked at the linux kernel source and it will set pollfd.revents=POLLHUP
> when the socket is closed, regardless of the value of  pollfd.events.  Not
> sure if every OS behaves this way, but maybe someone who knows other OS
> can provide some feedback.

Good to know Linux does that, so doing the proposed change on Linux would then 
certainly be safe.

> Ultimately, it would probably be better (faster) to only poll a single fd
> (the connection socket) and find some other way to detect if the server is
> shutting down,  but this change makes a huge difference in system load
> with lots of connections.

Well, other methods (pipes, signals) can be problematic for some systems in 
terms of portability, which is why I did not use any of those.  I think using 
POLLIN for systems where POLLHUP is otherwise not signalled should be fine -- 
and if needed we can probably even write a configure-test to detect what the 
situation is.

I'll have to look a bit more into this before changing the code though...

Happy hacking,

Christian
 
> -eivind
> ________________________________________
> From: address@hidden
> address@hidden On Behalf Of Eivind
> Sarto address@hidden Sent: Tuesday, March 22, 2011 5:16 PM
> To: address@hidden
> Subject: [libmicrohttpd] MHD_handle_connection() performance issue
> 
> I am using libmicrohttpd to serve a large number of video players that is
> using a "smooth" streaming format. That means the video stream is divided
> into separate chunks and that the client can decide to jump to a faster or
> slower encoding of the video stream on any chunk boundary.  The length of
> a video chunk is typically several seconds (<= 10 sec).
> (If you have 1000 clients all playing videos in 10 second chunks, you will
> have 100 requests/sec)
> 
> I wrote a client simulator so I can do some load testing without buying a
> whole bunch of client devices.
> 
> The server gets extremely busy as crank up the number of clients.  I am
> running the server with MHD_USE_THREAD_PER_CONMHDNECTION | MHD_USE_POLL
> The reason for this appears to be that MHD_handle_connection() will poll
> both the connection socket and the listen/server socket.  And every time
> there is a new connection, all the active connections will return from
> poll, because they are also polling the listen/server socket. It gets
> pretty awful as you try and crank up 1000 simulated streams.  The cpus are
> almost completely saturated with each connection/thread polling again and
> again, each time a new connection is made.
> 
> I made a debug version of  MHD_handle_connection() that only polls the
> connection socket, and I can run several 1000 video clients with almost no
> cpu load.
> But, each connection thread can no longer detect if the server is shutting
> down.
> 
> Any suggestion or feedback about what could be done to improve this
> situation?
> 
> Note: I am using the latest svn version with the fixes to MHD_poll() for
> MHD_USE_THREAD_PER_CONMHDNECTION.
> 
> -eivind