[Cerowrt-devel] Field Report: Switching to bql-31 at home

[Cerowrt-devel] Field Report: Switching to bql-31 at home

[Richard Brown]

I tried swapping my WRT54GL/DD-WRT with bql-31 over the weekend. It was a very rocky experience, and I switched back to DD-WRT so that others in my house could use the Internet. Here's what happened:

My home network is 192.168.2.0/24. It's configured to be different from my office (192.168.1.0/24). I also have a couple static IP addresses (printer, file server) at 192.168.2.13 & 192.168.2.23 at home.

I installed bql-31 and then used the sed commands to adjust the advertised subnets to match the addresses at home. I had to manually tweak the printer and file server addresses because they were wireless. I put them in the CeroWrt SSID, 192.168.2.64/27

Performance out to the Internet seemed to be fine, and maybe even snappy. I was not able to do methodical performance testing because several bad things happened:

1) All mDNS detection failed. My local printer (Brother MFC-655CW) is wireless. It was impossible to connect the Mac computers to the printer until I put them into the same SSID/subnet. Then mDNS worked, but required rebooting most equipment (Mac computers & printer) to get it going the first time.

2) I also tried installing mDNSResponder package, but that had no effect on the printer discovery. I didn't configure the package or troubleshoot this...

3) The worst problem was that the OSX computers and the printer all reverted to CeroWrt-guest SSID unbidden. This, of course, screwed up the mDNS, and I couldn't print. In one case, I connected my laptop to the CeroWrt SSID, put the computer to sleep, waited a minute or two, then awakened the computer to find it switched to CeroWrt-guest SSID.

4) I installed the snmpd package. This worked a bit, except that retrieving two interface data samples in quick succession (say, retrieving ifInOctets a few seconds apart) would show a rate of zero between the samples, even if lots of data was really flowing. 30-second samples seemed to work as expected. Does the interface not update the SNMP counters with every packet?

5) I installed fprobe to get NetFlow data. This seemed to be working fine, but was not extensively tested.

Observations:

a) I believe I read about troubles with CeroWrt wireless "dropping out". Could this be related to the switching of SSIDs that I observed? (I did not try Windows or Linux computers on wireless.)

b) mDNS seems to fail when it's routed. I believe there's a provision in the spec for a "mDNS reflector". I also found this link: http://groups.google.com/group/macenterprise/browse_thread/thread/0211bb0a02ef09b6 

c) SNMP counter data was not always reliable, since the counters didn't always seem to be updated.

d) I don't know how different this would be from using a recent OpenWrt on the WNDR3700v2.

What other information could I collect for further debugging? Thanks!

Rich

Re: [Cerowrt-devel] Field Report: Switching to bql-31 at home

[Richard Brown]

[-- Attachment #1: Type: text/plain, Size: 3508 bytes --]

Dave,

Thanks for the speedy responses. I should note that I had  a bunch of fun trying out bql-31: we all knew it wasn't ready for prime time, and now we know more of the symptoms :-) I look forward to taking another whack at this (probably this weekend).

I agree with everything, the note below adds detail where needed.

re: My connection. I'm using ADSL (Fairpoint) with ~2.8mbps down, 0.7 mbps up.

re: load on wireless. No, I did not do any organized load testing of the wireless network. I did download a few files, ran speedtest.net<http://speedtest.net>, etc. though. In that fairly informal test, CeroWrt and DD-WRT performed substantially the same.

re: mDNS over routed subnets. A stable avahi reflector really will make a difference. (I should say that I'm a big fan of mDNS. I cut my networking teeth on AppleTalk, whose Name Binding Protocol was the prototype for mDNS. At Dartmouth College, we had ~150 zones - one for each building. Also at Dartmouth, I wrote an AppleTalk NBP-to-IP translator that gave a stable AppleTalk name to an IP address, anticipating the dyndns.org<http://dyndns.org> facility...)

Then mDNS worked, but required rebooting most equipment (Mac computers & printer) to get it going the first time.

I think I default to an overly long dhcp server lease.

I'll be attentive next time I try this, to see if this is reproducible.

3) The worst problem was that the OSX computers and the printer all reverted to CeroWrt-guest SSID unbidden.

Now, THAT is weird.

Yeah. I will check this carefully with the next install.

4) I installed the snmpd package. This worked a bit, except that retrieving two interface data samples in quick succession (say, retrieving ifInOctets a few seconds apart) would show a rate of zero between the samples, even if lots of data was really flowing. 30-second samples seemed to work as expected. Does the interface not update the SNMP counters with every packet?

The interface does, but the snmp daemon probably has a polling interval set to high values. I'll look into it, however
most pollers do tend to run at 30 second to 5 minute intervals.

Interesting. I have not observed this with the snmp daemon of DD-WRT. Yes, I'd like to hear what you find. I may install the current OpenWrt as well to see if that exhibits the same symptom.

> I realize this is bucking the worldwide 'standard' of one network for all
> interfaces in the home, but I need it for the testing... and it is certainly interesting
> and revealing as
> to how much this is counter to people's expectations.
>
> IPv6 re-introduces this problem, too.

And it's forging new pathways. You're right that home routers may need to subnet for performance. My observations are all about making the "other stuff" work.

> How are you connecting to your office? VPN?

I fire up the VPN from my laptop into the office, not from my home network to the office. (No need for my wife and son to have that access.) In order for the PPTP software to work, the home and office subnets need to be different.

thx for giving it a shot, and giving me a grip on your expectations. I'm emphatically not a mac user...

As I said, I was happy to do this. I realize that the things I am reporting here are quite distinct from the goals of fixing bufferbloat. So I'm content simply to report them so that they're recorded. (At my company, our unofficial motto is, "We promise not to fix it if we don't know it's broke.")

Best regards,

Rich

[-- Attachment #2: Type: text/html, Size: 4886 bytes --]