| initcwnd adjustment using RHEL5 kernels? [message #42897] |
Wed, 15 June 2011 07:46  |
aquarapid
Messages: 4
Registered: January 2010
|
Junior Member |
|
|
When using a vanilla RHEL5 kernel, you can adjust the TCP initial congestion window (to 10, instead of the default 3, assuming a normal mss) by setting the initcwnd option on the default route. Something like:
ip route change default via x.x.x.x dev eth0 initcwnd 10
I can confirm that this works using the vanilla CentOS el5 kernel 2.6.18-238.9.1.el5.
I have been trying to make it work using the OpenVZ RHEL5 kernels (specifically using el5.028stab091.1, but I've also tried el5.028stab089.1). Eventually I want to use it inside a container, but for now I'm just testing using CT0 (one step at a time...).
However, I'm not having any luck at all. Setting the initcwnd via "ip route" succeeds, but I am still seeing the normal TCP slow start behavior when fetching files via HTTP from the machine. I've checked all the obvious problems, e.g.:
* the default TCP send buffer size (set via net.ipv4.tcp_wmem) is indeed >= 10 times mss
* make sure to flush the route cache between tests.
* Restart the (HTTP) server application between tests for good measure.
Can anyone confirm that they have raised initcwnd in VE0 (or even better, inside a container) and actually seen it work?
J
|
|
|
|
|
|
| Re: initcwnd adjustment using RHEL5 kernels? [message #42902 is a reply to message #42897] |
Wed, 15 June 2011 10:38  |
aquarapid
Messages: 4
Registered: January 2010
|
Junior Member |
|
|
Nevermind, I found the problem, which is quite interesting.
I was setting the sysctl net.ipv4.tcp_no_metrics_save=1 in my openvz sysctl.conf, which seems to kill the initcwnd adjustment and cause fallback to defaults.
As soon as I tracked this down and reset it back to the default value of 0, the initcwnd adjustment works as expected.
J
|
|
|
|
OpenVZ Forum
Members
Search
Help
Register
Login
Home
