ndt-users list created

[Phil Reese <>]

Rich,

The server is built upon Ubuntu 5.04, Debian, kernel 2.6.10.

I've not tuned the stack as I've not seen suggestions for how and what 
to tweak. Happy to tweak something, but don't know what.

Just googled and find that Pittsburgh Supercomputing Center has the 
following web page about Linux TCP tuning: 
http://www.psc.edu/networking/projects/tcptune/#Linux, however it 
suggests that web100 systems do much of the tuning themselves.

Here are the values for my current TCP stack:
preese@netspeed:/proc/sys/net/ipv4$
 more tcp_mem
24576   32768   49152
preese@netspeed:/proc/sys/net/ipv4$
 more tcp_rmem
4096    87380   174760
preese@netspeed:/proc/sys/net/ipv4$
 more tcp_wmem
4096    16384   131072

I'm at the office now and don't have a copy of the 'More Details' screen 
from my home computer so I'll have to gather that info tonight.

With Craig's comments on similar speeds between Stanford's NDT and DSL 
Reports, while my speeds from home continue to be very different, I have 
to wonder if there are network issues between Comcast and Stanford's 
commodity net connection?

Mike's suggestion on Seg Faults rings true too, however I don't see any 
of them in any of the logs in /var/log/*.

Its looking like there is some isolation of this problem to one or 
several specific providers.  Does this make sense to anyone?

Thanks for all your analysis!

Phil


On 9/13/2005 5:53 AM, Richard Carlson wrote:
> Hi Phil;
>
> What version of Linux are you using?  Have you tuned the stack 
> (/proc/sys/net/ipv4/tcp_*mem)?  Also, look at the "More Details" window 
> and let me know what CurCwnd, CurRwinRcvd, Duration, and Sndbuf values are.
>
> Off the top of my head, either the Linux stack is untuned and the server 
> is limiting the throughput, or the NDT has a bug and the test is running 
> longer than it should, (and skewing the results).
>
> Rich
> At 07:55 PM 9/12/2005, Phil Reese wrote:
> > All,
> >
> > From notes from several users and my own experience, I've found that 
> > my NDT server seems to greatly underrate inbound (server to client) 
> > bandwidth. However, from my campus machine, the speeds (inbound and 
> > outbound) seem about right and consistent with tests just after I 
> > deployed the new server.
> >
> > As a data point, I've run the NDT test and gotten results and then run 
> > a DSL reports speed test.  The DSL report site usually gives me the 
> > speed I believe I have at my home via cable modem, while NDT reports 
> > drastically different results.
> >
> > For example:
> >
> > NDT
> > running 10s outbound test (client to server) . . . . . 757Kb/s
> > running 10s inbound test (server to client) . . . . . . 2.78Mb/s
> >
> > DSL Reports
> > Upload    768K
> > Download  7.34M
> >
> > Also, NDT provides an extra line suggesting that setting my received 
> > buffer to a different value would give better results.  Originally, I 
> > was getting results via NDT very similar to the DSL numbers and no 
> > additional commentary on the receiver buffer size.
> >
> > Any ideas what might have happened or is happening with my server?
> >
> > I've rebooted and stopped and restarted the services several times. 
> > I'm using 5.3.3e version of the frontend.
> >
> > Any help would be appreciated!
> >
> > Phil Reese
> > Stanford
>
> ------------------------------------
>
>
>
> Richard A. Carlson                              e-mail: 
>
> Network Engineer                                phone:  (734) 352-7043
> Internet2                                       fax:    (734) 913-4255
> 1000 Oakbrook Dr; Suite 300
> Ann Arbor,  MI  48104