Is it right way to test between two modes?

Is it right way to test between two modes?

[이재홍]

[-- Attachment #1: Type: text/plain, Size: 712 bytes --]

Hello,
I'm really sorry to ask a lot... but I want to make sure that what I did is
right or not

I compared two modes
1. default poll mode
2. interrupt mode (In my understand.. it is kind of an event driven mode in
dpdk)

[Test 1] - poll mode
 $ sudo ./examples/dpdk-l3fwd -l 1-3 -n 4 -- -p 0x3 --config="(0,0,1)"
* power: about 36W*

[Test 2] - event driven mode
 $ sudo ./examples/dpdk-l3fwd-power -l 1-3 -n 4 -- -p 0x03
--config="(0,0,1)" --interrupt-only
* power: about 12W*


When Idle situation, Test1 uses a core 100% and Test2 doesn't use a core a
lot (almost 0%), so I think I can say "In idle mode, Test2 Mode(event
driven) reduces power by 1/3 compared to poll mode." Is it correct??

BR,
Jaehong Lee

[-- Attachment #2: Type: text/html, Size: 1026 bytes --]

Re: Is it right way to test between two modes?

[Stephen Hemminger]

On Thu, 11 May 2023 13:47:23 +0900
이재홍 <ljh890322@gmail.com> wrote:

> Hello,
> I'm really sorry to ask a lot... but I want to make sure that what I did is
> right or not
> 
> I compared two modes
> 1. default poll mode
> 2. interrupt mode (In my understand.. it is kind of an event driven mode in
> dpdk)
> 
> [Test 1] - poll mode
>  $ sudo ./examples/dpdk-l3fwd -l 1-3 -n 4 -- -p 0x3 --config="(0,0,1)"
> * power: about 36W*
> 
> [Test 2] - event driven mode
>  $ sudo ./examples/dpdk-l3fwd-power -l 1-3 -n 4 -- -p 0x03
> --config="(0,0,1)" --interrupt-only
> * power: about 12W*
> 
> 
> When Idle situation, Test1 uses a core 100% and Test2 doesn't use a core a
> lot (almost 0%), so I think I can say "In idle mode, Test2 Mode(event
> driven) reduces power by 1/3 compared to poll mode." Is it correct??
> 
> BR,
> Jaehong Lee

A couple of other notes.
1. Generating real world traffic pattern is important. Testing with something
like a full line rate test of UDP is not real life. Ideally a pattern of multiple
real machines to see the self-correlated TCP bursts, etc.

2. Measuring actual power would require something attached to the power supply.
The CPU load is not that great a final measure.  One option would be to use
something like powertop which can look at some of the kernel internals.