Very strange measurements

[Brecht Vermeulen]

Hi,

last year (August) I did some measurements with Smartbits equipment on a
AMD K6-550 MHz PC with Tulip cards (DFE-570 4 port card) and click 1.0.6
(2.2.14 kernel).

Now I repeated some tests with a AMD K6-500 MHz and an AMD Athlon 1GHz
and get some strange results with the Athlon platform :

* with click snapshot from cvs (30 May 01, but similar results with a
snapshot from 4-07) :
64 byte packets, load from 10 Mbit (10% of the link) - 100 Mbit in steps
of 10 Mbit :

Frameloss (in %) :
10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
0.0000	0.0000	0.1346	2.2245	3.6076	3.9371	7.1374	0.0000	0.0000	0.5883
0.0000	0.0000	0.1346	2.2245	3.6076	3.9371	7.1374	0.0000	0.0000	0.5883

very strange is the frame loss around 60-70 Mbit and again a loss of 0
at 90-100 Mbit and it is reproducable !

Avg Latency (in us) :

4219.3	3357.3	3264.5	2564.7	2384.2	901.2	1123.0	89.5	82.5	2792.1
4219.3	3357.3	3264.5	2564.7	2384.2	901.2	1123.0	89.5	82.5	2792.1

again very strange numbers, they decrease with a minimum at 80-90 Mbit
(at 100% there can be some artefacts)

* however, I took click 1.0.6 (tested also click 1.1 and later, but
these gave the same results) with a 2.2.14 kernel for the same platform
:

10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
0.0000	0.0000	0.0060	0.0000	0.0000	0.0000	0.0004	0.0008	0.0010	0.6055
0.0000	0.0000	0.0060	0.0000	0.0000	0.0000	0.0004	0.0008	0.0010	0.6055

everywhere 0 loss !!! with same hardware

Avg latency :
315.3	292.4	279.5	230.7	236.4	198.7	237.8	125.8	138.0	2755.7
315.3	292.4	279.5	230.7	236.4	198.7	237.8	125.8	138.0	2755.7

yes, 10 times smaller ! but still decreasing with bigger load.

* and now it comes, the same network card in a K6-500 MHz

frame loss :

10%	20%	30%	40%	50%	60%	70%	80%	90%	100%
0.0000	0.0000	0.0000	0.0003	0.0074	0.0019	0.0371	0.0034	0.0034	4.1549
0.0000	0.0000	0.0000	0.0003	0.0074	0.0019	0.0371	0.0034	0.0034	4.1549

Avg latency :
18.0	17.7	17.7	18.3	21.1	22.4	23.0	23.6	24.7	2887.9
18.0	17.7	17.7	18.3	21.1	22.4	23.0	23.6	24.7	2887.9

another 20 times faster and logical increase with bigger load !!!

for this K6 platform, the results with recent cvs snapshots and click
1.0.6 were the same.

Does anyone has an explanation for this behaviour (and what was the big
chance between click 1.0.6 and later versions ?) or do we have to burn
our GHz processors :-) ?

The delayshaper element works perfect (well, on the K6 platform :-) ),
e.g. a delay of 1 ms is *exactly* 1 ms. A delay of 20us is *exactly* 20
us !!! However I think it would be more logical to make the element
accept us values as configuration (as everything in the element is based
on us and you can give ms also , but now it is not possible to give us
delays without changes) (the changes take approx. 3 changing lines :-)
).

best regards,
Brecht