dslreports logo
uniqs
12

telcodad
MVM
join:2011-09-16
Lincroft, NJ

telcodad to Hostage

MVM

to Hostage

Re: Comcast's new speed test doesn't do upload

said by Hostage:

said by winterfl:

... At 1000/35, you are using 100% of your upstream just for the TCP ACK packets (with some even being dropped!). Good luck trying to upload anything at all while downloading at 1000Mbps.

If Comcast were being reasonable about uploads, upload speed would be at a minimum 8% of download speed.

Comcast's 150/5 plan is a little worse than that. Nice 30x slower upload 👌.

I think the worst ratio though is the "Extreme Pro" 400/10 plan, with a 40x slower upload, and therefore only 2.5% of download speed!

Darknessfall
Premium Member
join:2012-08-17
Motorola MG8725
Cisco DPC3008
Asus RT-N66

4 edits

Darknessfall

Premium Member

said by telcodad:

I think the worst ratio though is the "Extreme Pro" 400/10 plan, with a 40x slower upload, and therefore only 2.5% of download speed!

Does anyone know how much upload is needed per Mbps of download of traffic due to ack packets or whatever? Isn't it riding it close?

Edit: One person here said that when downloading 840 Mbps, it used 21 Mbps upload (40x) »Re: why are upload/download ratios so horrible in Canada?

Wouldn't a light bandwidth upload on 400/10 basically kill the download speeds?
peteg
join:2016-05-08
Bellevue, WA

1 edit

peteg

Member

I believe this is every-other, in most cases. This would imply about half the 35 (about 18Mbps) is being "eaten by ACKs", in the "high use" scenario, but...
I do recall something about recent Docsis also suppressing "large blocks of ACKs", based on them all being neutral (no issue).

How you figure out how much "suppression" is actually going on is probably a "Wireshark exercise", I bet, your ISP is never going to publish that data ;-]
winterfl
join:2010-06-09
Mount Dora, FL

4 edits

winterfl to Darknessfall

Member

to Darknessfall
said by Darknessfall:

said by telcodad:

I think the worst ratio though is the "Extreme Pro" 400/10 plan, with a 40x slower upload, and therefore only 2.5% of download speed!

Does anyone know how much upload is needed per Mbps of download of traffic due to ack packets or whatever? Isn't it riding it close?

Edit: One person here said that when downloading 840 Mbps, it used 21 Mbps upload (40x) »Re: why are upload/download ratios so horrible in Canada?

Wouldn't a light bandwidth upload on 400/10 basically kill the download speeds?

So, I actually did some extensive testing on this.

You end up with some strange efficiency that works in Comcast's favor here ... Some amount of packet loss on ACK's is acceptable because each ACK packet basically means that every thing prior to that byte counter has been received ... Normally, ACK's are sent for every-other packet, but the sending side is also sending a window of packets at a time, (ie, dozens, or even hundreds of packets may be 'in flight' before an ACK is received for what was sent).

Because of this dynamic, significant amounts of the ACK packets can be dropped while full speed is achieved, as long as there is no packet loss in the downstream direction, in which case the limitations of the upstream are more problematic.

The absolute bare minimum I've found to be able to achieve a 1Gbps download when all other conditions are ideal is 7.38 Mbps upstream capacity.

Keep in mind that this number would make the service really, really, really annoying to use, as it would mean that any interactive service would be completely unusable during a download (ie: a game would flat out be impossible), and also any packet loss at all would drastically reduce performance. Packet loss always hurts, but this would make it much more extreme.

At 400Mbps, that would mean that technically they could get away with a minimum of 3Mbps, if they didn't care about providing service which was usable for anything interactive, or working in anything less than perfectly ideal situations.

Also, any upload /at all/ would drastically cut download speeds in this sort of scenario. Like, even sending an email with a jpeg attached would hurt your download speeds, if the upstream was cut this low.

In order to have something that is able to handle even moderately less than ideal scenarios, 24 Mbps upstream would again be the 'bare minimum' for 1Gbps downstream. On a 400 Mbps service, that would work out to about 9.6Mbps. So, that's probably where comcast arrived at 400/10.

Franken
join:2016-02-26

Franken

Member

said by winterfl:

The absolute bare minimum I've found to be able to achieve a 1Gbps download when all other conditions are ideal is 7.38 Mbps upstream capacity.

At what end did you measure it?
winterfl
join:2010-06-09
Mount Dora, FL

winterfl

Member

I tested by using a packet shaper to reduce the amount of bandwidth available to ack's, and continued reducing the amount available until download speeds suffered.

Keep in mind, this is a lab test assuming all other conditions are ideal, which, in the real world, they aren't.

Franken
join:2016-02-26

Franken

Member

So you didn't measure it at the other end, at the sender? You need to measure it there to know how many ACKs the modem is actually transmitting (because of ACK suppression).
winterfl
join:2010-06-09
Mount Dora, FL

winterfl

Member

I just tested, and the amount of ACK's sent from my desktop is exactly equal to the amount of ACK's received on my server in a datacenter.

No evidence of ACK Suppression here. Not sure if that's something turned on/off by the CMTS, or if it's a limitation of the MB8600 I have.

Franken
join:2016-02-26

Franken

Member

So you ran a network sniffer at your server? Just to understand your setup
winterfl
join:2010-06-09
Mount Dora, FL

winterfl

Member

Yes.

Tcptrace at both ends.