TCP Tuning issue
 

Hello,

I have a question about tuning TCP parameters for a WAN link in order to facilitate high throughput, basically keeping the pipe full. I am only changing parameters on the receiver side which is running Red Hat 5.3. Kernel - 2.6.18-53.el5.

With the default values for everything - net.ipv4.tcp_wmem, net.ipv4.tcp_rmem, net.ipv4.tcp_mem, etc .the behavior that I see is that the sender sends 2 packets with size 1448 bytes and then seems to wait for an ACK from the receiver before it sends another 2 packets of the same size.

I then change the above mentioned variables to the foll:
#Increase TCP maximum buffer size
net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
#Increase Linux autotuning TCP buffer limits
net.ipv4.tcp_rmem = 4096 87380 16777216
net.ipv4.tcp_wmem = 4096 65536 16777216
net.ipv4.tcp_mem = 2097152 2097152 16777216

With the above the behavior is exactly the same as the first case. With the bigger buffers, shouldn't I see more data in flight? What am I missing here and how can I increase my data in flight? Just for the heck of it, I also tried:

net.ipv4.tcp_rmem = 2097152 2097152 16777216
net.ipv4.tcp_wmem = 2097152 2097152 16777216

with everything else the same as case 2 but no change. Windows scaling is enabled.

I would appreciate any help with this.

Thanks!
-p

0) Linux (kernel) auto-tunes {r,w}mem on boot according to available RAM, 1) receiver-side modification of the stack can AFAIK only indirectly influence sender behaviour by way of return traffic and 2) tuning the stack doesn't take into account bottlenecks like those of the sender-side (I/O, CPU, sysctl, tc, firewall), router (packet limiting?) and Internet "weather". It would make sense to first determine sender maximum throughput, the realistic (as opposed to advertised) throughput of your pipe, how routing affects throughput and from that glean realistic upper and lower traffic volume bounds.