Investigation into the Effects of e1000 version 5.2.20 Module Parameters on TCP Traffic

Date November 2003

Goal

The purpose of this investigation is to quantify and qualify the effects of the various modules parameters within the e1000 5.2.20 driver on TCP traffic. The introduction of the 5.2.20 driver brings along a couple of new features:

• InterruptThrottleRate: this enables an absolute maximum rate of interrupts that are to be generated by the driver. This value should reduce cpu utilisation - although it may result in lower performance.
• NAPI compatibility: When compiled in, this should enable the kernel to switch between poll and listen modes. At high speeds, this should reduce the cpu utilisation.

Method

A series of tests will be performed; back to back and across the MBNG network. It is preferential to try tcp tests across longer bandwidth delay products (MBNG is only 750kB), however, due to hardware constraints, this was not possible.

The following variables will be investigated, both with and without the NAPI enabled. The InterruptThrottleRate was set to auto (1).

• txint with default paramters on sender
• txint with default parameters on recv
• rxint with default parameterson sender
• rxint with default parameterson recv
• txdescriptor with default parameters on sender
• txdescriptor with default parameters on recv
• rxdescriptor with default parameters on sender
• rxdescriptor with default parameters on recv
• {tx|rx}AbsIntDelay?

Setup

The tests were conducted with modified 2.4.20 kernels. The kernel was altAIMD version 0.2. Tests were conducted with the vanilla kernel with ABC disabled and moderate_cwnd() off.

Txqueuelens were set to a value greater than anticipated (2000) for all tests.

Results

Without NAPI

With NAPI

Conclusion

{blank}

Wed, 19 November, 2003 16:18