ndt-users list created

[Bill Abbott <>]

I've tried tcp send and receive (max) buffers from 64k up to 8 MB, no 
improvement.

Smaller buffers are more consistent, larger buffers tend to go up then 
hit zero, then climb again.

The current settings are:

$ cat /proc/sys/net/ipv4/tcp_rmem
4096    87380   8174760
$ cat /proc/sys/net/ipv4/tcp_wmem
4096    16384   8131072

I've also modified a bunch of other stuff, but it didn't make a 
difference so I set it back.  ECN, /proc/sys/net/core/rmem_max and 
wmem_max, netdev_max_backlog, etc.




Matthew J Zekauskas wrote:
> To follow up to myself...
>
> * The Internet2 servers are all gigE, and all 9000-byte MTU capable.
>
> * Another thing you could do is tune for the bottleneck instead of 
> gigabit..
>
>    150 Mbps = 150*1000*1000 bits * 0.07 ms RTT = 10500000 bits
>      /8 = 1312500 bytes /(1024*1024) = 1.26 megabytes.
>
> or
>
>    100 Mbps is 850K
>
> and see if that improves things.  It may be that something can't handle 
> even that much back-to-back when they are sent at 1000Mbps.
>
> There are also bwctl (<http://e2epi.internet2.edu/bwctl/>) servers at 
> Internet2 router nodes, for a different way to split the problem 
> (essentially run iperf to the core).
>
> --Matt