Feedback → speedtest gigabit testing/results/debugging

feedback topic just for gigabit+ connections using »/speedtest

Why is the speed less than (insert my other local favoured test here)?

1. Be careful you are comparing a test of internet speed, and not testing to your ISP core or very close to it. People tend to gravitate to the test that gives them the fastest possible reading. That's fine but realise what you are testing. Your last mile can go X megabits per second, can your ISP deliver that much speed and low latency too, to the net?



2. If you are comparing Flash, browser performance, coupled with potentially other issues such as disk speed or anti-virus products. Browser performance is a real problem.

Observed desktop browser performance on a 4+ gigabit connection, late 2016:

Chrome 53 all: good on download (disk i/o), good on upload (http recommended)
Firefox 49 OSX: poor on download, poor on upload, VERY poor on https
Firefox 49 Linux: good on download, good on upload, VERY poor on https
Safari 10 OSX: poor on download (excessive memory use), good on upload, network connection errors on https
Opera 39 OSX: very good on download (disk i/o), very good on upload, good on https

SUMMARY: Towards and beyond a gigabit poor https performance becomes a big issue. You must switch to http otherwise decryption slows the browser. No browser even without decryption is anywhere near as efficient as a command line tool such as netperf or iperf3. Opera can achieve 3-4 gigabit on a macbook pro under OSX, however. Most browsers use all of two CPU cores when reading or writing over a gigabit and are thus cpu-limited even if they are not disk speed limited.

* memory use can be excessive. Some browsers are more efficient than others. With very high speed connections memory use by the browser can blow out to 1 2 or 3 gigabytes and doesn't always get reclaimed after the test finishes.

* where there is disk i/o, your disk has to write at least as fast as the download speed of your connection. So a gigabit connection writes at ~120mb/second.

Some bookmarks maybe helpful:

* »rolande.wordpress.com/20 ··· -part-2/
A good description of tuning for OSX for maximum performance at and above 1 gigabit. Goes into the importance of memory buffers, and sizing MSS optimally. Fairly up-to-date.

* »www.duckware.com/blog/ho ··· dex.html
An examination of the current limitations of Windows 7+ and high speed TCP. Fairly up-to-date.

* »calomel.org/network_perf ··· nce.html
A good view of the PC hardware necessary for going faster than 1gig. OpenBSD tuning tips

* »fasterdata.es.net/host-tuning/
Some collected articles on tuning various host OSs.

Things to look at, if you want to go faster (updated mid 2016)

0. Try Opera. Seriously.

1. Turn off any "real time" anti-virus products and firewalls! Mcafee is a problem. You can disable "real time" protection for a time period. This is the first thing to try if you have issues. The same goes for any browser extensions that you suspect could be filtering or processing data.

2. Check whether CPU or Disk is saturated. Use whatever tool you have to watch CPU user and kernel time during a test. See if one or more cores are saturated. Check whether there is excessive disk activity.

3. If you are using Chrome, try the "experimental" option in preferences. It does not use any disk i/o. Unfortunately it uses a lot of CPU and the top speed is about a gigabit. But since it avoids disk, this may be sufficient.

4. Use the server "Auto Select" button to concentrate on the nearest servers.
If you are curious use the "By Bandwidth" button to see what server speeds are like for single transfers, one by one. This will also reveal the lower bound on the number of streams you need.

5. Use http not https -- the speed test will function in either non-encrypted or encrypted pages but for encrypted, all the data-transfers must be encrypted as well and this can cause excessive CPU use.

6. Tune your TCP stack, use the fasterdata.es.net site for tips. If there is a very close speed test server, or the speed you want to reach is a gigabit or less, this may not be necessary.

Tracing route to v-162-248-92-123.unman-vds.internap-chicago.nfoservers.com [162
.248.92.123]
over a maximum of 30 hops:

1 1 ms 1 ms 1 ms yong-10a-c6504e-01-vlan1998.tcom.purdue.edu [10.
160.1.1]
2 88 ms 12 ms 2 ms itns-service-vss-01-vlan2100.tcom.purdue.edu [17
2.28.249.1]
3 1 ms 1 ms 5 ms 172.28.246.0
4 1 ms 1 ms 2 ms tel-210-c9006-01-te0-0-0-2.tcom.purdue.edu [172.
28.252.85]
5 2 ms 2 ms 2 ms indiana-gigapop-ctc-internet-151.tcom.purdue.edu
[192.5.40.82]
6 7 ms 7 ms 7 ms 149.165.183.90
7 10 ms 7 ms 15 ms equinix-ix.ord1.us.voxel.net [206.223.119.107]
8 7 ms 7 ms 7 ms internap.ord1.us.voxel.net [208.122.29.22]
9 7 ms 7 ms 7 ms border10.po1-bbnet1.chg.pnap.net [64.94.32.22]
10 7 ms 7 ms 7 ms inap-b10.e5.router.chicago.nfoservers.com [64.74
.97.254]
11 7 ms 7 ms 7 ms v-162-248-92-123.unman-vds.internap-chicago.nfos
ervers.com [162.248.92.123]

and for the speedtest server used *I believe*
Tracing route to mail.tzulo.com [208.77.16.10]
over a maximum of 30 hops:

1 1 ms 1 ms 1 ms yong-10a-c6504e-01-vlan1998.tcom.purdue.edu [10.
160.1.1]
2 1 ms 1 ms 1 ms itns-service-vss-01-vlan2100.tcom.purdue.edu [17
2.28.249.1]
3 8 ms 2 ms 1 ms 172.28.246.0
4 39 ms 39 ms 39 ms tel-210-c9006-01-te0-0-0-2.tcom.purdue.edu [172.
28.252.85]
5 38 ms 38 ms 89 ms indiana-gigapop-ctc-internet-151.tcom.purdue.edu
[192.5.40.82]
6 44 ms 43 ms 42 ms et-10-0-0.211.chic0.tr-cps.internet2.edu [64.57.
21.13]
7 51 ms 71 ms 43 ms equinix-chicago.r1.chi1.us.as5580.net [206.223.1
19.45]
8 44 ms 44 ms 47 ms eth1-4.r2.chi1.us.as5580.net [78.152.35.241]
9 43 ms 43 ms 43 ms 78.152.33.246
10 8 ms 7 ms 7 ms mail.tzulo.com [208.77.16.10]

Ran another test: »[Fiber Speed test: 460.82/616.28 13 ms]

Ok before I make an option to actually search for the fastest routes, instead of picking randomly by lowest latency, can you select just Chicago and New York
and select the number of download streams to 24, and upload to 12 then do a run?

Also if you can let me know the CPU and memory of whatever machines you have tried so far ?

Another user was getting 300 mbit with a mac that had a core 2 duo, and went to a mac-mini (not a paragon of speed) and doubled their download to 600 because I think of the i5 processor.

What is this message and why is my system considered?

0.2s Forcing lo-fi mode due to slow CPU or linux desktop

Current system being tested - OS Name Microsoft Windows 7 Professional

Processor AMD FX(tm)-6100 Six-Core Processor, 3300 Mhz, 3 Core(s), 6 Logical Processor(s)

Installed Physical Memory (RAM) 16.0 GB

there is a quick javascript benchmark done involving a sort of floating point loop and below a threshold, the cpu is classified as slow for the display of the graphs during the test.
It is inexact, but your browser "scores" 8000 vs a more typical score of 20 to 120k for desktop.

What might cause the browser to be so high?
I have never heard of anything like this

slow, you mean?
AMD cpu isnt very good, ie11 has a slower javascript engine than other browsers.
Anyway it doesnt change the result of the test, thats the main thing.

i figure the guy(s) who designed the speed test would know a thing or two about this

what's the best packet scheduler and queuing discipline?

im using fiber fios internet 75/75 with an Asus AC56U from the ONT and a Jan 2016 build of DD_WRT

HTB or HFSC scheduling

and

SFQ or CODEL or FC_CODEL queuing

most articles are from like 2006 and the earliest is 2011, id like to know how those affect the 2016 internet of data streaming and online gaming and whatnot

I don't know the answer however if you can find the bufferbloat mailing list archives, you can find information there and/or join and ask them?

I'm having issues with the test on my Gigabit connection, I only get 880 or so, I get 940 on Flash tests.

I got an i3-6100.

Did you try more than one browser?
check % cpu during the test?
check lo-fi in the preferences to make it easier for the browser?

I tested with Edge, and got under 500 Mbps lol

I use about 60% CPU, and enabling or disabling HTTPS doesn't make a difference.

lo-fi doesn't really change anything

Can you check a disk activity graph?

Unfortunately due to the design browsers have they insist on writing everything to their cache files, even if its not to be cached. If your disk subsystem is not up to gigabit write speed, or at least not up to it after the layers of i/o functions up to the browser, then the limit will be the disk saturation point.

For example to get 4 or 5 gigabit on my iMac I have to create a ram disk and point the browser cache directory to it in preferences. Otherwise the disk is the bottleneck.

I got a SATA 3 SSD, and it's at 30% or so

If you want to waste more time I recommend seeing if you can put the browser cache path onto a temporary ram drive. Although the result you get isn't that far below maximum, it may still be the cause. (Flash doesn't write anything to disk as it downloads).

Also you could try Firefox just to see if it matches Chrome or goes faster (or slower).

The other reason could just be that all but one of the "nearest servers" are in USA not Canada, and 60ms away or more.

yeah i'll do that, I got plenty of time to waste lol

The other thing you could try is increase download streams from 32 to 40 I seem to remember 32 isn't a hard limit. I doubt it will change anything but if it does it would show the bottleneck is server distance and selection, not browser or cpu etc.
Your result seems to be one of the fastest dots in the bell.ca grid btw.

I haven't been able to change the cache to a RAM drive, but I tested 40 streams, and it seems to slow it down.

I also tested with Edge and I get pretty much the same results

By connecting two computers locally over a 2gig cable (but using a ram disk for browser cache) I've found Chrome, Firefox and Safari are all capable of downloading at full speed.

Either the flash server you are using is so local that it is not striking the same bottle neck, or there is a browser/cache/cpu issue that is holding the html5 version back. If you add more streams and it doesn't go faster it indicates the bottleneck is not the server side but is either on the PC, or on the gateway out of bell.ca that is not an issue with the flash test if its server is in-network.

There is a command line version you can try that eliminates disk and browser as a factor. I'll send you the link.

okay, so the command line version uses about 20% CPU, while the browser one uses 50 or so.

The results vary wildly, and I got once over 930 Mbit/s on task manager, but the tool reported under 900.

It rarely goes above 700, in the few tests that I did, and the Flash test i'm using reports 930, but uses 70% CPU.

Also, the fastest result in the grid has a CPU that's rated at 22392, and mine is at 56933, so I guess it's fast enough...

Where is the ookla test server located that gives you 930 consistently using flash?

Montreal, I'm using Fibrenoire's server.

They may have a better, or direct, peer with Bell.ca that is making the gigabit product look optimal. If I had to guess.

Could be. Thanks!

Okay, so, random speedtest, 904 Mbits, with HTTPS. Can you see if something is different about that run? I did another one without HTTPS, and got only like 730.

I think you won't be able to get a rock solid 940 every run, clearly whatever you get depends a bit on some external factors: to either the server network (on my side), or your PC and home connection on your side. It isn't https vs http more like just unpredictability.

You're also not very placed in Quebec regarding servers that are nearby. There is one nearby one at teksavvy, and the next nearest is new york. The rest are far.

I feel that your speedtest is very inconsistant compared to the ookla ones. Might only be wind blowing the photons away lol

I feel that your ISP is not consistently able to get you gigabit speeds from "the Internet". The results for other gigabit users on other gigabit ISPs are stable.
If the speed is variable then it indicates there is congestion. You are using 32 streams from half a dozen servers, but when you add more it makes no improvement which is a sign that there is a capacity issue on the routes. For all I know, all of the test server network might be for you via one route.
Go to server preferences and just select pulling from teksavvy, Google and Amazon servers only. What speed do you get?

I don't think there's *that* much congestion, if any. My neighborhood isn't a new one or anything, and it's got a lot of older people. It's also quite a small city (60k people)

Might also be an issue with the ISP provided router, which I haven't yet been able to completely remove.

I'll try what you said tomorrow tho, thanks!