[Cerowrt-devel] Problems testing sqm

[Sebastian Moeller]

Hi David,


On Oct 24, 2015, at 18:34 , David P. Reed <dpreed at reed.com> wrote:

> Not trying to haggle.

	Sorry, I was a bit to grumpy for unrelated reasons.

> Just pointing out that this test configuration has a very short RTT. maybe too short for our SQM to adjust to.

	That could be, but I believe people have tested fq_codel and sqm with similar setups and generally got dozens of milliseconds induced delay, not multiple seconds. So sure sqm might not for the best thing but it should deliver a reasonable compromise. Now, I believe Toke has a test bed where he can vary the transmission delay so he might know already whether sqm has issues with 1GE lans.

Best Regards
	Sebastian

> 
> On Oct 24, 2015, Sebastian Moeller <moeller0 at gmx.de> wrote:
> Hi David,
> 
> On Oct 24, 2015, at 00:53 , David P. Reed <dpreed at reed.com> wrote:
> 
> In particular, the DUT should probably have no more than 2 packets of outbound queueing given the very small RTT. 2xRTT is the most buffering you want in the loop.
> 
> Let’s not haggle about the precise amount of queueing we deem acceptable, as long as we all agree that >= 2 seconds is simply not acceptable ;) (the default sqm will approximately limit the latency under load increase (LULI) to roughly twice the target or typically 10 ms; note that this LULI only applies to unrelated flows). The exact number of queued packets seems to correlate with the beefiness of the DUT, the beefier the fewer packets should work, wimpier devices might need to batch some processing up, resulting in higher LULI…
> 
> Best Regards
> Sebastian
> 
> 
> On Oct 23, 2015, Richard Smith <smithbone at gmail.com> wrote:
> On 10/23/2015 02:41 PM, Michael Richardson wrote:
> Richard Smith <smithbone at gmail.com> wrote:
> My test setup:
> 
> Laptop<--1000BaseT-->DUT<--1000baseT-->Server
> 
> So, given that the DUT is the only real constraint in the network, what
> do you expect to see from this setup?
> 
> Given that the probably DUT can't forward at Gb/s, and it certainly can't
> shape anything, it's gonna drop packets, and it's probably gonna drop them in
> Rx, having overrun the Rx-queue (so tail-drop). If there is too much ram
> (bufferbloated), then you'll see different results...
> 
> Setting ingress/egress to 10Mbit/s I expected to see the speed 
> measurements bounce around those limits with the ping times staying in 
> the low double digits of ms. What I saw however, was the data rates 
> going well past 10Mbit limit and pings up to 2000 ms.
> 
> This is what I've seen in prior rrul testing using a the 50/10 cable 
> link at our office and my 25(ish)/6 link at my apartment and a well 
> connected server on the net. That however was using QoS and not SQM.
> 
> Its that a reasonable expectation?
> 
> -- Sent with K-@ Mail - the evolution of emailing.
> 
> Cerowrt-devel mailing list
> Cerowrt-devel at lists.bufferbloat.net
> https://lists.bufferbloat.net/listinfo/cerowrt-devel
> 
> 
> -- Sent with K-@ Mail - the evolution of emailing.