ndt-users list created

[Dale Blount <>]

On Wed, 2006-02-15 at 10:40 -0500, Richard Carlson wrote:
> Hi Dale;
> 
> I don't recall if I replied to this earlier but here it is again.
> 
> I'm seeing this problem more and more and I am currently trying to find a 
> real fix for this problem.  Here's what's happening.
> 
> The NDT server starts sending data to the remote client.  It enters a 
> simple send loop and pumps data into the network for 10 seconds.  It then 
> exits the loop and sends the final results over to the client.
> 
> The problem is, that when in this loop it is possible for the OS to 
> transfer data to the TCP stack, faster than the TCP stack pump it out into 
> the network.  This results in a large standing queue, visible with the 
> netstat -nat command).  So at the end of 10 seconds the code stops pumping 
> more data but the client keeps reading until the queue empties.  Note the 
> Web100 Duration variable has a value of 34,893,925 microseconds, or almost 
> 35 seconds.
> 
> One temporary step is to limit the max buffer space (tcp_wmax) to something 
> in the 512 KB to 1MB range.  This will keep the queue from building up too 
> much, but it's really just a band-aid until I can figure out how to monitor 
> the queue length to prevent such large queues in the first place.
> 

Rich,

I can't find a tcp_wmax setting, but here is what I have set:

net.core.wmem_max = 131072
net.core.rmem_max = 131072
net.ipv4.tcp_wmem = 4096 16384 131072
net.ipv4.tcp_rmem = 4096 16384 131072

Upload always works OK, but on anything but the lan, download is right
around 75k.  It doesn't really matter if it's set to 128kb/512kb/2Mb,
it's always 70-80kb (both on a 5Mbps upload cable modem and a 768kbps
upload dsl link, both 3 hops from the ndt server).

The old server that this is replacing is still around, and speedtests to
it work just fine.  Could the newer hardware alone be causing this whole
problem?  I've tried the sysctl settings from the old server with the
same results.

> If anyone has any suggestions on how to do this, please let me know.
> 

Couldn't the client be adjusted to stop reading after 10 seconds?  It
could then report the data transferred so far.

Thanks,

Dale