Kitz Forum

Internet => General Internet => Topic started by: Weaver on December 21, 2016, 12:08:39 AM

Title: Time variation in TCP throughput & latency - differences between bonded lines
Post by: Weaver on December 21, 2016, 12:08:39 AM

The screenshot at
    https://flic.kr/p/PMbR19
shows a long download from Amazon using my three lines IP-bonded together as recorded by Andrews & Arnold’s clueless server. I'm trying to understand certain aspects of what I'm seeing.

Firstly the lines aren't showing the same behaviour, and I wonder why that might be. (The line speeds are not identical, that's one thing.)

The dark green line at the top shows the download rate which is 2 Mbps per modem. The lines are numbered 1, 3 and 4, for hysterical reasons. Line 4, the lowest, shows a flat throughput trace, whereas line 3 in the middle has periodic dips, in a regular pattern which is not the same as that shown on line 1. Presumably something to do with AA’s scheduler.

Secondly, I wonder what's going on with the latency picture - the green and blue at the bottom of the picture. Green above is max latency, bright blue below it is the average, according to the key on the right. The pictures are not remotely the same between the lines.

There is so much I don't understand. Could anyone enlighten me?

Title: Re: Time variation in TCP throughput & latency - differences between bonded lines
Post by: huwwatkins on December 21, 2016, 10:51:04 AM

If Line 4 is running flat out (as the slowest line) and the others are at say 95% then that would possibly explain the difference in latency.

Title: Re: Time variation in TCP throughput & latency - differences between bonded lines
Post by: aesmith on December 21, 2016, 12:08:35 PM

At a guess something to do with the scheduler.   The top line shows intermittant spikes in utilisation above the general 2meg line, the second shows intermittant dips below.  Almost as if it sometimes managed to squeeze an extra packet into the first line, and every now and then had to miss one out on the second.   Alternatively could be jitter in the measurements, with the sample plotted on the graph sometimes representing more than one second, sometimes less.

Would be interesting to see the same view when the line(s) are fully loaded with multiple downloads rather than one TCP stream.

Title: Re: Time variation in TCP throughput & latency - differences between bonded lines
Post by: Weaver on December 21, 2016, 02:23:09 PM

I could compare the multiple TCP case by setting up some Netflix downloads, as I think their app runs four downloads in parallel, so it's a reasonable guess that that might involve four TCP connections. I can check by doing a packet capture.

I wonder what is going on with the latency pictures?