Incast Experiment

[ChengjunJia]

Hello, I want to repeat the incast experiment in the PowerTCP paper.
The result of 10-1 incast fits well with the paper, but when I add more flows (I tried the 96-1 incast), the result is different from the paper (the queue is about 100KB-200KB after coverage).

The flows I use are:

97
0 16 3 10000 1000000000000 0.13
32 16 3 10001 1000000000000 0.15
33 16 3 10033 1000000000000 0.15
...
127 16 3 10127 1000000000000 0.15

Is there any problem with my experiment? In the paper, you say

in addition to the 10 : 1 incast, the 256th server sends a query request (§4.1) to all the other 255 servers which then respond at the same time, creating a 255:1 incast.

How can I create the experiment?

[vamsiDT]

Hi,

The 255:1 incast scenario is composed of (4 long flows if I remember correctly) and a query request that generates 255 flows (you would need 255 short flows approx 4K bytes in size). Launching 255 long flows (building up 255BDPs of buffer) is an extreme case where no congestion control algorithm cannot do any better, given the small buffer size.

Hope it is clear.

[ChengjunJia]

I got it. Thanks very much :)

[HeRaNO]

Hi @vamsiDT !

I'm also trying to reproduce the result of HPCC in paper. I use the following topology:

257 1 1 256
256
0 256 25000000000.0 1us 0
1 256 25000000000.0 1us 0
2 256 25000000000.0 1us 0
...
255 256 25000000000.0 1us 0

and following flow file:

266
0 16 3 10000 1000000000000 0.13
1 16 3 10001 1000000000000 0.15
2 16 3 10002 1000000000000 0.15
3 16 3 10003 1000000000000 0.15
4 16 3 10004 1000000000000 0.15
5 16 3 10005 1000000000000 0.15
6 16 3 10006 1000000000000 0.15
7 16 3 10007 1000000000000 0.15
8 16 3 10008 1000000000000 0.15
9 16 3 10009 1000000000000 0.15
10 16 3 10010 1000000000000 0.15
0 16 3 10300 4000 0.15
1 16 3 10301 4000 0.15
2 16 3 10302 4000 0.15
3 16 3 10303 4000 0.15
...
255 16 3 10555 4000 0.15

But I found that the result is still different from the paper. My result shows as follows.

Could you please tell me whether the topology or the flows are right for reproducing? If something is wrong, would you please to point them out? Thanks!