I suppose we need a real world example. E.g. how would one configure http to be prioritised over ftp using this interface?

I think you could just specify the ports and assign them to 2 priority queues.

In reply to a post by BatBoy:

I think you could just specify the ports and assign them to 2 priority queues.

Yeah, Source/Dest Port is configurable, so I suppose where that is a known value it can be used in QoS. In the case of Robocopy, I could experiment with Dest Port 445 (the PCs on the LAN seem to be communicating via link-local IPv6).

In reply to a post by Oliver341:

That sounds useful. I saw the definition "Specifies the inter-packet gap to free bandwidth on slow lines" which I thought would not be applicable to 100 mbit LAN, but I will certainly give it a go.

any link running at 100% can be considered a slow link

What you're all describing are the effects of bufferbloat. There are too many, unnecessarily large buffers in far too many network devices between you and the destination. TCP by its very definition, given sufficient CPU power, will fill up any buffer given enough time. Once the buffer is full then your packet, whatever size, whatever priority will have to wait its turn in the queue as data drains away down the slow pipe.

The trick is to manage the queue/s and not quite fill the pipe, thus latency (which is the measure you are using to measure how 'sticky' your network connection is) can be controlled. Once latency is managed, interactive/time critical packets can be interwoven amongst the longer term 'but hog all the bandwidth' data flows. The huge 'elephant' flows as they are known can be managed by dropping packets to tell the other end to slow down.

This allows space for the 'ants' (DNS, ICMP, VOIP, TCP/SYN/ACK - small usually single packet but desperately important) and 'mice' (Short, small transfers) to do their work and give the interactive (ie low latency) experience you actually required.

Google for cerowrt, Dave Taht, Jim Gettys, bufferbloat, 'tcp ants mice elephants', fq_codel.

https://gettys.wordpress.com/bufferbloat-faq/

The cerowrt bufferbloat project fed a lot of information back into the mainline Linux kernel, and latest versions of openwrt with 'SQM Scripts' (for fq_codel queue management) do extremely well in managing latency and perceived performance of bloated edge of internet routers.

If that doesn't scare you then the talk http://cyber.law.harvard.edu/events/luncheon/2014/06... certainly should!

Friends don't let friends run factory firmware.

Ignore the red bits and anything pre 12:00 but at 1830ish I started a 2GB 'one drive' upload along with a 1GB Ubuntu image download (that's the 6min duration slightly higher latency spike) Interactive performance during this time has been pretty much unaffected. This is on a 40/10 Sky IPv4 link with a HG612 modem with TP-Link Archer C7 AC1750 running Openwrt CC latest trunk (as of 12:00 today ish) and SQM scripts - SQM limits are set to 38000 down & 9500 up respectively.

Oh and then one of the machines upstairs started doing some sort of automatic download...all pretty much unnoticeable.

My Broadband Ping

Edited by deleted (Tue 14-Apr-15 20:05:28)

If the only tool you have is a hammer, then every problem becomes a nail.

In reply to a post by Oliver341:

That sounds useful. I saw the definition "Specifies the inter-packet gap to free bandwidth on slow lines" which I thought would not be applicable to 100 mbit LAN, but I will certainly give it a go.

I wonder if your router is using the CPU to copy packets to/from the LAN ports and running out of CPU ?

Perhaps try putting a cheap 5port separate switch in front of the router (ie, all clients on router into this switch, and one link to the router) as that should ensure LAN traffic stays on the switch and WAN traffic goes to the router.

I've been using OpenWrt with fq_codel for the past couple of years now with great success (see here for full details of my config including the multiple wireless access points question i noticed you asked about recently on the CeroWrt list).

I don't tend to do big uploads to services like Dropbox because i host everything on my FreeNAS server but because of this there's almost always uploads and downloads happening on my connection. Thanks to fq_codel no matter what i do my connection remains extremely responsive (even with the netperf-wrapper tests the additional latency stays within 10ms) usually within 3-5ms of additional latency.

People just don't seem to realise how bad a problem bufferbloat is. It's absolutely everywhere and the fix is simple; use a router that supports fq_codel.

In reply to a post by jchamier:

Perhaps try putting a cheap 5port separate switch in front of the router (ie, all clients on router into this switch, and one link to the router) as that should ensure LAN traffic stays on the switch and WAN traffic goes to the router.

Yeah, could do.

I've had a play with some of the Huawei QoS settings. Certainly it's very easy to see the "Policer" in action, since it's a straight speed throttle. Creating a Policer with CIR and PIR set to 100 limits the speed to 100 Kbps. Assigning port 22 (SFTP) to it throttles it to 10 KBps and it's verifiable in Filezilla.

However, putting port 445 into the Policer and running Robocopy had no effect on the slowdown, which makes me think the QoS on this router is purely for WAN traffic. But even so, it's a nice feature and seems to work well for WAN traffic.

The IPG option in Robocopy is a workable solution. It obviously slows down the transfer considerably when a sizeable gap is specified, but since it's generally being run as a background operation, it's not time critical.

When I enabled QoS on my router (using tomatousb) to prioritize ack's my max speed dropped to under 50mbps due to the fact the router cpu could not handle scanning all the packets for my line speed, so the problem QoS has with faster connections is you need quite a beefy cpu. Also it tends to add bufferbloat and for it to be effective you generally need to have some bandwidth in reserve.

funny enough the hg612's first firmware had QoS on by default which took 10% of the upload speed.

Edited by Chrysalis (Thu 16-Apr-15 11:13:44)