Segment length Throughput

asked 2018-11-04 09:40:29 +0000

dazzpowder gravatar image


We have a network spanning application when running on one AWS instance type caused network file copies to be really slow throughput around 30MB/s. the Avg packet size was around 3kB segment length avg about 8KB and there were many retransmisions and dup acks . I put the slowness down to the app somehow chopping up the data into loads more packets causing many more ACKS and poor performance. When the instance type was changed for one with better network performance we saw the throughput practically double now at 75MB/s but now the Avg packet size was 18KB so this made some sense as to why it was quicker as a lot more data was going on the wire requiring less ACk's. Although confused by why the segment length AVG was now only 170B

What I'm confused me even more was I changed the AWS instance type again for one with even better network performance using AWS propitiatory driver and the throughput increased again, but I cant see how as the AVG packet size is now 14KB, segment Length was about 1350

How is it possible that a smaller AVG packet size can now be quicker than larger AVG sizer? Any help as to what I can look at in terms of bottleneck? I know AWS imposes some kind of network bandwidth throttling but I cant see it reflected in the captures.

Any help making sense of this much appreciated

edit retag flag offensive close merge delete


Can you clarify your terms a little - what do you mean by packet size and segment length? A segment is the TCP payload, so if you say it's 1350, the packet (which would be IP, which again is the Ethernet payload) should be in the range of 1390 bytes. A frame (Ethernet plus playload) would be 1408, including FCS. So 14KB doesn't match that at all.

My assumption is you're talking about the TCP MSS being 1350, and you see frames of 14KB and more?

Jasper gravatar imageJasper ( 2018-11-04 10:38:19 +0000 )edit

Thanks for the response.

three captures in terms of throughput all copying over same size data Bad Max frame.len 55090 frame.len < 100 = 42821 Avg B/s 33 Average pps 10940.3

Good Max frame.len 1514 frame.len < 100 = 5978 Avg B/s 68 Average pps 47662.8

Best Max frame.len 65226 frame.len < 100 = 7529 Avg B/s 75 Average pps 4101.7

How does a host with way smaller but consistent frame length have almost the same throughput as a host with higher frame length. By changing the network interface type I'm able to pushe more data through I would expect frame lengths to be the same just greater pps. If there is some kind of throttling/capping or even cos on the interface how can this be seen in a capture?

dazzpowder gravatar imagedazzpowder ( 2018-11-04 16:25:41 +0000 )edit