Matthew,
Firstly, thanks for the help :-)
Are you watching for an EV_WRITE event while you have nothing to
write?
Nope -- my static Write() method is only called by libEvent, and it
(the Write() method) doesn't loop. In fact, here it is in its'
entirety:
void CSocketTestModule::Write( int iDescriptor, short sEvent, void
*pArg )
{
// pArg is the struct event ...
struct event *pEvent = reinterpret_cast< struct event * > ( pArg );
// write as much data as we can to this client ...
CClient *pClient = mapOClients[ iDescriptor ];
if ( pClient )
{
iWrites ++;
int iSendLimit;
if ( iSize - pClient->GetBufferPosition() > pClient-
>GetLowWatermark() )
{
iSendLimit = pClient->GetLowWatermark();
}
else
{
iSendLimit = iSize - pClient->GetBufferPosition();
}
int iBytesSent = send( pClient->GetDescriptor(),
pClient->GetBufferPosition() + pBuffer,
iSendLimit,
0 );
if ( -1 != iBytesSent )
{
iTotalBytesWritten += iBytesSent;
pClient->IncBufferPosition( iBytesSent );
}
// if the client is still around and we've more data to
send, add the event again
if ( iSize > pClient->GetBufferPosition() && iBytesSent != 0 )
{
event_add( pEvent, NULL );
}
}
else
{
std::cout << "no client instance for this FD!?!?\n";
}
}
Are you attempting to send more data in a single send() than the
socket's
SO_SNDBUF and SO_SNDLOWAT allow?
Nope -- I only send SO_SNDLOWAT bytes (or less, if i'm at the end of
the clip, for instance) in each send(). I also tried sending
SO_SNDBUF bytes in each send(), but neither helped -- CPU usage is
way high (90%+), while only streaming a 6mb/s MPEG2 clip to 1 VLC
client on another machine.
I'm not understanding why the CPU is so high ... i'm really anxious
to find out :-)
I'm still learning libEvent, so maybe there is something I can do to
fix this. I've a very basic C++ server based on libEvent, and i'm
having the server stream 1 MPEG2 PS (6-8mbps) file to 1 client over
HTTP. The client is VLC. While this works, the CPU usage is very
high e.g. 90+% on both my PowerBook (running OSX 10.4.6) and a (much
faster) Debian Sarge box running on a 2.4ghz P4. This CPU usage is
the server only -- the VLC client is on another machine.
The program sits inside event_dispatch() and doesn't come out until
the client disconnects. Furthermore, if I use VLC to stream the same
file over RTP, it only uses 2-4% of the CPU, not 90+%.
I'm very new to libEvent, but is there something I can do to make the
CPU usage of my libEvent-based server more inline with what VLC
uses? If I do an strace (on Linux) or a ktrace (on OSX), I see the
same calls -- epoll_ctl(), a send() which returns EAGAIN (since I
made the socket non-blocking), and that's about it. Is there a call
I can make into libEvent before I call event_dispatch() that will
make libEvent not spin so heavily on the select() / kqueue / epoll()?
Regards,
John
Falling You - exploring the beauty of voice and sound
http://www.fallingyou.com
_______________________________________________
Libevent-users mailing list
Libevent-users@monkey.org
http://monkey.org/mailman/listinfo/libevent-users
