Experts please explain the bandwidth overhead to me!

Hi geeks.

Not an issue, just something I would like to understand. Have been testing backups to cloud which of course will benefit from higher UL speeds.

Test scenario:
Test PC uses a dedicated line that has a theoretical maximum of 6.2 MBps UL speed, or 50 Mbps. Only test PC is connected to this line, all other network usage denied or disabled. The line is only used for this test.

To measure UL speed, I start uploading to a cloud server. Test backup size 30 GB. A counter starts measuring the speed when it hits 5.8 MBps, marking then actual speed every 20 seconds.

This shows averages after about 20 tests:



Horizontal values show elapsed time in minutes. Variations in all tests max 10% to any direction, the fact seems to be that after about 3 minutes the UL speed sets in around 2 MBps and stays there. Results are similar with XP Pro, Windows 7 and Windows 8 Enterprise editions so it does not look as an OS related thing.

I am very interested to hear opinions in this, why over 60% overhead on a dedicated line?

Kari

Fiberoptics 100/50 Mbps line, during the test dedicated to this one PC alone. Cisco router that can for sure handle the speeds, third party speed tests giving constantly over 90% of promised speeds. Downloads keep constant 12 MBps +- 0.5 Mbps.

Receiving end also Cisco routers, even faster line and a dedicated server only working for this test.

This is clear: cause of overhead is not linespeed or inability to handle it.

Kari

I've found that one thing that severely impacts transfer speed is the file size - the bigger the file, the faster the transfer. I guess it has to do with the number of data exchanges required at the start/end of a file.

It's one reason that I like SkyDrive more and more - if you download multiple files from a folder, it autozips them prior to transfer, so you not only get less to move, it moves it faster :) (Dunno what happens with upload though).

Good point Noel, the file size I mean. It's just that this system I am testing is supposed to move system images over the network so the files are huge.

Interesting results from tests run at night between 3 AM and 4:45 AM, big difference compared to daytime results. Only difference is the time of the day:



Average overhead 32%, compared to earlier tests 60%.

Thinking about the shape of your first graph, it reminds me a lot of the sort of bottleneck you may get from a buffer that's slow to flush.
- initial high speed as the buffer fills, then quick drop-off to a steady state as the buffer empties and is refilled.
You may want to check what all the links in the line use as buffers and see if there's some multi-tasking reserving anything.

Very good idea! Thanks, Noel.