Packet loss in IP/UDP Stack?

Packet loss in IP/UDP Stack?

[Martin Rauh]

Hi folks,

It seems that we are loosing packets in the UDP stack.
Somebody might think that this is not astonishing, but
the packets are not lost at the network layer. They seem to get
lost in the IP/UDP Layer of the receiving box.

We have got the following configuration:
Two linux boxes (P4, 733 MHz, 256 MB RAM, kernel 2.4.0)
are directly connected with two syskonnect sk98xx
gigabit ethernet cards. We are sending a file from one host to the other

with UDP. About half of the file is lost in the receiving application.
But according to the statistics in /proc/net/dev no errors (fifo, frame,
dropped...)
occured in the network layer. Even the transmitted and received
data at the network layer (tx-bytes, -packets) are identical
to the amount of the transfered file (plus network overhead).

Doas anybody know where the loss occurs? Is there a loss at the
network layer or in the higher protocol layers?

Many Thanks,

Martin Rauh

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: Packet loss in IP/UDP Stack?

[Andi Kleen]

On Fri, Jan 26, 2001 at 07:01:12PM +0100, Martin Rauh wrote:
> Hi folks,
> 
> It seems that we are loosing packets in the UDP stack.
> Somebody might think that this is not astonishing, but
> the packets are not lost at the network layer. They seem to get
> lost in the IP/UDP Layer of the receiving box.
> 
> We have got the following configuration:
> Two linux boxes (P4, 733 MHz, 256 MB RAM, kernel 2.4.0)
> are directly connected with two syskonnect sk98xx
> gigabit ethernet cards. We are sending a file from one host to the other
> 
> with UDP. About half of the file is lost in the receiving application.
> But according to the statistics in /proc/net/dev no errors (fifo, frame,
> dropped...)
> occured in the network layer. Even the transmitted and received
> data at the network layer (tx-bytes, -packets) are identical
> to the amount of the transfered file (plus network overhead).
> 
> Doas anybody know where the loss occurs? Is there a loss at the
> network layer or in the higher protocol layers?
 
UDP by default has 64K buffer and drops packets when it overflows.
You can check that using netstat -s. 
You can increase it use the SO_{SND,RCV}BUF socket options and/or the 
net/core/[rw]mem_* sysctls.



-Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: Packet loss in IP/UDP Stack?

[Martin Rauh]

The UDP packet loss went away by setting the rmem_max and rmem_default
in /proc/sys/net/core on the receiving linux box to about 80 percent of the
size of the
transfered file (~50MB).

Many thanks for the fast help!

greetings,

Martin Rauh


Andi Kleen wrote:

> On Fri, Jan 26, 2001 at 07:01:12PM +0100, Martin Rauh wrote:
> > Hi folks,
> >
> > It seems that we are loosing packets in the UDP stack.
> > Somebody might think that this is not astonishing, but
> > the packets are not lost at the network layer. They seem to get
> > lost in the IP/UDP Layer of the receiving box.
> >
> > We have got the following configuration:
> > Two linux boxes (P4, 733 MHz, 256 MB RAM, kernel 2.4.0)
> > are directly connected with two syskonnect sk98xx
> > gigabit ethernet cards. We are sending a file from one host to the other
> >
> > with UDP. About half of the file is lost in the receiving application.
> > But according to the statistics in /proc/net/dev no errors (fifo, frame,
> > dropped...)
> > occured in the network layer. Even the transmitted and received
> > data at the network layer (tx-bytes, -packets) are identical
> > to the amount of the transfered file (plus network overhead).
> >
> > Doas anybody know where the loss occurs? Is there a loss at the
> > network layer or in the higher protocol layers?
>
> UDP by default has 64K buffer and drops packets when it overflows.
> You can check that using netstat -s.
> You can increase it use the SO_{SND,RCV}BUF socket options and/or the
> net/core/[rw]mem_* sysctls.
>
> -Andi
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> Please read the FAQ at http://www.tux.org/lkml/

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/