Hi,
I have a requirement to handle TCP connections directly, bypassing
the Linux stack. To do this, I have blocked all packets to the one port I'm
using with iptables and send/receive on a PF_PACKET socket.
This works well, but I am currently having a performance issue -
despite the modest hardware, I had expected to achieve greater than the 4000
packets a second that I am currently achieving (40 byte packets.) A similar
program running on NT achieves 25kpps.
The hardware is an ordinary 700MHz Athlon system with a 3c905C
ethernet card. It's doing nothing else, and is on a switched network with a
low level of broadcasts. I'm currently using a RedHat 2.4.18-19.8.0 kernel.
For measuring the packet rate, the code looks broadly like this:
if((s=socket(PF_PACKET, SOCK_DGRAM, htons(ETH_P_IP)))<0) {
[...]
len=recvfrom(s,&buf,MAX_LEN,0,(struct sockaddr *)&saddr_in,&saddr_size);
[...]
while(alrm) {
id++;
n=sendto(s,iph_out,40+data_len_out,0,(struct sockaddr
*)&saddr_in,saddr_size);
}
(data_len_out is 0.)
Is 4000 packets a second reasonable on hardware like this? I found
an old message in an archive that considered 20kpps for sendto() on a raw
socket was slow. I had certainly been expecting more.
Is SOCK_DGRAM significantly slower than SOCK_RAW for PF_PACKET?
Would an upgrade to a 2.5 kernel help? Where is the bottleneck likely to be?
I'd rather avoid a kernel module if possible, but if it's the only
way to get the performance I need, I'll consider it.
So, if anyone has any ideas, they'd be much appreciated.
Many thanks,
Richard
-
: send the line "unsubscribe linux-net" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
