On Oct 26, 8:04 am, Andreas Roth wrote:

> i'm using epoll and aio with TCP sockets on a Ubuntu 7.10 AMD64 system. I
> have one thread which waits for epoll events. The epoll thread informs me
> when new data is available (EPOLLIN) and i issue an aio read request
> (IO_CMD_PREAD). But under certain circumstances the epoll thread
> continously informs me about new data available to read.
> To my knowledge the submitted read request should eliminate further EPOLLIN
> notifications until new data from the network is ready. Am I wrong? Have
> anyone experienced the same problem?

You are expecting an asynchronous request to have a synchronous
effect. It does not. An asynchronous request may take effect at some
unspecified future time, out of synch with your epoll loop.

Are you sockets set non-blocking or blocking?

DS

David Schwartz wrote:

> On Oct 26, 8:04 am, Andreas Roth wrote:
>
>> i'm using epoll and aio with TCP sockets on a Ubuntu 7.10 AMD64 system. I
>> have one thread which waits for epoll events. The epoll thread informs me
>> when new data is available (EPOLLIN) and i issue an aio read request
>> (IO_CMD_PREAD). But under certain circumstances the epoll thread
>> continously informs me about new data available to read.
>> To my knowledge the submitted read request should eliminate further
>> EPOLLIN notifications until new data from the network is ready. Am I
>> wrong? Have anyone experienced the same problem?
>
> You are expecting an asynchronous request to have a synchronous
> effect. It does not. An asynchronous request may take effect at some
> unspecified future time, out of synch with your epoll loop.
To avoid starting the same receive operation again i need to ignore the
EPOLLIN notifications. To ignore them is easy, but the epoll loop is still
active and is spinning. I could remove the socket from the epoll after a
recv operation has been started, but this is quite expensive and in most
cases not necessary.
>
> Are you sockets set non-blocking or blocking?

All sockets are non-blocking.
>
> DS

Thanks four your help.
--
Andreas Roth

On Oct 26, 11:39 pm, Andreas Roth wrote:

> To avoid starting the same receive operation again i need to ignore the
> EPOLLIN notifications. To ignore them is easy, but the epoll loop is still
> active and is spinning. I could remove the socket from the epoll after a
> recv operation has been started, but this is quite expensive and in most
> cases not necessary.

Why not just perform the 'read' synchronously? The socket is non-
blocking, so the 'read' operation won't block. As soon as you finish
reading, but before you process any data, allow the 'epoll' thread to
make forward progress.

I usually do this with a counter. When 'epoll' returns, as I handle
each read or write job, I increment a counter. As I complete the
socket operation (but no other operations needed), I decrement the
counter. When the counter reaches zero, I can try to discover new
sockets.

DS

On Sat, 27 Oct 2007 08:39:38 +0200, Andreas Roth wrote:
> David Schwartz wrote:
>> On Oct 26, 8:04 am, Andreas Roth wrote:
>>> i'm using epoll and aio with TCP sockets on a Ubuntu 7.10 AMD64 system. I
>>> have one thread which waits for epoll events. The epoll thread informs me
>>> when new data is available (EPOLLIN) and i issue an aio read request
>> You are expecting an asynchronous request to have a synchronous
>> effect. It does not. An asynchronous request may take effect at some
>> unspecified future time, out of synch with your epoll loop.
> To avoid starting the same receive operation again i need to ignore the
> EPOLLIN notifications.

Would EPOLLET (edge triggered) help?

But I still don't really see the point why you are trying to use
asynchronous IO at all. BTW, is the Linux aio implementation actually
asynchronous or does it use a synchronous read request behind the scenes?


Christof

--
http://cmeerw.org sip:cmeerw at cmeerw.org
mailto:cmeerw at cmeerw.org xmpp:cmeerw at cmeerw.org