As a result of upgrading a number of servers from CentOS 4 to CentOS 5, we have noticed that multicast latency has increased. I realize that CentOS 4 uses the e1000 driver for intel nics, and CentOS 5 uses e1000e, however after further investigation and testing I have found that there are earlier version of the driver that deliver better performance than others.
I conducted a 'bounce' test as one server sends data, and the second receives and sends it right back.
After running multiple tests following the driver adjustments we made, I found some interesting and noteworthy patterns:
Driver Version CentOS Driver Version CentOS Avg RTT
< 1.2.7 4 or 5 <1.2.7 4 or 5 75ms
< 1.2.7 4 or 5 1.3.10 5 90-100ms
<1.2.7 4 or 5 1.9.5 5 90-100ms
1.9.5 5 1.9.5 5 140ms
- The version of CentOS itself does not seem to matter. In addition, the old e1000 driver used in CentOS 4 appears to perform similarly to the 1.2.7 version of e1000e.
- Upgrading from 1.3.10 to 1.9.5 did not produce different results either way ? making the case that whatever has changed has remained consistent.
- After upgrading from 1.2.7 to 1.9.5, there were sizeable performance decreases ? especially when communicating with another interface running 1.9.5. The performance of 1.9.5 communicating with a < 1.2.7 driver was consistent that of 1.3.10.
Does anyone have any ideas or insight as to where along the line Intel may have changed something in their drivers? I've looked through the documentation that intel has packaged with their drivers, however do not see any difference in default settings, new features, etc.
Any input is appreciated.
Issues related to configuring your network
4 posts • Page 1 of 1