In reply to a post by dogzilla:

It's usually expected that you won't get as much speed in a single thread so that's a difficult problem to prove.

Not if you're on a proper ISP it isn't.

http://www.thinkbroadband.com/speedtest/button/14769...

That's actually quite bad for a PlusNet graph. I suspect the x6 jumping around is more down to my laptop and Flash though. I don't have a more recent test to show because I've disabled Flash and now use the HTML tester.

Edited by Andrue (Mon 21-Nov-16 10:10:28)

Or a difference in how provider is handling HTTP versus a generic TCP socket protocol on the flash test.

Real world across numerous networks and appreciable distances, it has always been the case that multi-threaded beats single threaded.

Firstly I'm not a network engineer (so that might explain it...),but my day job involves access to decent connectivity, I've been in or around datacentres for the past 15 years and have spent lots of time testing and or playing with some big connections as well as troubleshooting performance issues.

I don't remember ever coming across a real world situation where single threaded was faster when trying to max out a connection. I mean yeah if you are doing an iperf between two servers in the same rack and you've got decent switches with decent TCP window size then yeah i guess the single threaded is going to be identical if not faster (less overhead to move the same amount of data)


Here is a couple of tests from a server with 10 gig connection:

http://www.thinkbroadband.com/speedtest/results.html...
http://www.speedtest.net/my-result/5817969981

Quite the variation in results..... And I've tried this from about 10 different sites.

IMO the best method of judging your connection speed is a multi-threaded download from a VARIETY of sources. E.g. JD's auto speed tester - configured to do a multi-threaded download from a variety of different sources so that I'm not affected by slowdowns by a particular route or a particular server.

Some VM customers are affected by a very obvious single thread download speed problem. I'm one of them and used to be able to achieve 200Mb on single thread downloads, but over the last year it has dropped to around 50Mb.

Meanwhile speedtest.net continues to show 200Mb, but these speeds are simply no longer achievable on real world downloads. It makes a mockery of VM's claims such as "Download a movie in x minutes with VIVID 200" when the only way you can reach 200Mb is on multithread speedtests.

http://www.thinkbroadband.com/speedtest/results.html...

Edited by deleted (Tue 22-Nov-16 11:52:52)

In reply to a post by dogzilla:

Real world across numerous networks and appreciable distances, it has always been the case that multi-threaded beats single threaded.

That's not in dispute but the degree of difference is.

In reply to a post by dogzilla:

I mean yeah if you are doing an iperf between two servers in the same rack and you've got decent switches with decent TCP window size then yeah i guess the single threaded is going to be identical if not faster (less overhead to move the same amount of data)

Can't see how TCP window size impacts a local connection where round trip should be < 500uS.

In reply to a post by dogzilla:

Here is a couple of tests from a server with 10 gig connection:

http://www.thinkbroadband.com/speedtest/results.html...
http://www.speedtest.net/my-result/5817969981

Quite the variation in results..... And I've tried this from about 10 different sites.

Those results (even the speedtest.net) look appalling for a 10Gb/s connection. Or are you saying the server had a 10Gb card fitted?

In reply to a post by dogzilla:

IMO the best method of judging your connection speed is a multi-threaded download from a VARIETY of sources. E.g. JD's auto speed tester - configured to do a multi-threaded download from a variety of different sources so that I'm not affected by slowdowns by a particular route or a particular server.

There is no real dispute that you can measure your connection speed using multiple concurrent socket connections and this will probably lead to a higher figure but this really does not represent a true reflection of the real word speeds.

The point of this thread, however, is to look at a comparison between DOCSIS and VDSL2 based connections. To do this appropriately I would suggest that the test method needs to be representative of real world use. In my view this means comparison of single thread speeds on the two systems.

Here's my VDSL2 result at 12:54 today.

A typical speed test graph on a VDSL2 line

I am able to push 400mbit/sec over the Atlantic from France never mind just to London from a uk address, to latency and RWIN will not be the reason for VM customers assuming they not using a really old OS like XP which has no auto tuning for tcp buffers. anything vista and newer utilises auto tuning and likewise linux and FreeBSD have had it for years.

In reply to a post by RobertoS:

In reply to a post by dogzilla:

It's usually expected that you won't get as much speed in a single thread so that's a difficult problem to prove.

Try this tester. The dark green is single stream and the light green six-stream.

Uncongested they will be the same or thereabouts.

LOL mine is the other way around :/

http://www.thinkbroadband.com/speedtest/results.html...

Yup = yummy

http://beta.speedtest.net/result/5818850336

I have absolutely no faith in the TBB tester. I just applied the test and got 120 Mb on the green and 116 Mb on the yellow. The single thread test actually does better than the multi-thread!. TTB has not reflected my actual broadband availability for a couple of year on either test.

With Speedtest.net and the DSLreports test (both multi-thread) I get I invariable get > 310Mb. I just got 262Mb on a single thread test in Testmy.net. I even get over 150Mb on US based servers.

Edited by deleted (Tue 22-Nov-16 18:23:07)

I've just dug around and found your test I believe. http://www.thinkbroadband.com/speedtest/results.html...

Single thread average 119.6 and multiple thread 146 Mbps.

DSL Reports can be 32 threads from various locations so any congestion is almost certainly bypassed, and the sampling in how speedtest.net behaves favours a higher result plus may be picking a data source on Virgin Media network. Ookla even claims that Sky has a 300 Mbps average speed service in London, so either their analysis is very wrong or some very odd tests being counted by them.

We can crank up thread counts and drop single thread if people do not trust it, but experience has shown it does reflect real world for many people.