dslreports logo
 
    All Forums Hot Topics Gallery
spc
Search similar:


uniqs
758

pflog
Bueller? Bueller?
MVM
join:2001-09-01
El Dorado Hills, CA

pflog

MVM

Powerboost and FreeBSD

Click for full size
This thread got me testing out my 22/5 business connection.

What I'm seeing is really quite annoying. I am hoping it is something I can fix on the FreeBSD side, or perhaps it's a misconfigured shaper on Comcast's side, I don't know.

Attached is an example throughput graph of what I'm seeing for a single connection. It ramps up to well over 8 MB/s with powerboost for a few seconds, then plummets to below 100 KB/s, then bursts again and finally levels off at 22 Mbps or so.

What can I do to try to avoid this, beyond setting my own traffic shaping/queuing to keep it from power boosting so high then crapping out due to the inherent packet loss following the power boost?

If anyone at Comcast with some clout is reading this, please consider allowing those of us who don't need/want power boost to disable it on our accounts. I'll take consistency over these erratic speeds any day. Though I imagine I'm in the minority.

SpaethCo
Digital Plumber
MVM
join:2001-04-21
Minneapolis, MN

SpaethCo

MVM

You might want to try implementing a different scaling algorithm. Cubic (used on newer Linux kernels) seems to deliver quite favorable results.

»caia.swin.edu.au/urp/new ··· ols.html

There may be other sources for drivers, but these seem to be current and derived from trusted sources.

pflog
Bueller? Bueller?
MVM
join:2001-09-01
El Dorado Hills, CA

pflog

MVM

Thanks for the link I'll give these a shot when I have some time. For now, I'm going to enable ipfw and dummynet so I can throttle the iface to 22 Mbps. I already have a script I used before to limit a particular UID to 10 Mbps when I had FiOS, so a little tweaking and I should have something. Long term, I'd like to get this modified TCP stack working, though, then pending good results I'll start campaigning for getting one or more of these merged into -CURRENT

nmbdude
@rr.com

nmbdude to pflog

Anon

to pflog
what port are you using to generate those graphs?

Also, as you already hinted at-- using pf on freebsd to qos yourself to 22Mbit/sec, you could solve the problem yourself, and not *that* complex...
nmbdude

nmbdude to pflog

Anon

to pflog
just so you know, you'll get much better fairness / latency with RED instead of Fair Queueing, so pf with altq will probably be more desirable than ipfw with dummynet.

Dummynet is a bit more versatile for some uses, but it's probably unlikely that it will be more useful for your needs.

pflog
Bueller? Bueller?
MVM
join:2001-09-01
El Dorado Hills, CA

pflog to SpaethCo

MVM

to SpaethCo
Click for full size
So I have a mail thread going with one of the guys who is working on that new tcp stack on FreeBSD, and he suggested I try setting net.inet.tcp.reass.maxqlen higher.

I increased it until I no longer saw net.inet.tcp.reass.overflows increasing when downloading from the xmission mirror. I currently have it set to 740 (default is 48 !) and I'm no longer seeing the drop off and net.inet.tcp.reass.overflows isn't increasing anymore.

I need to find out if setting net.inet.tcp.reass.maxqlen this high might have other undesired effects, but it certainly does help with this particular download test.

New tcp throughput graph attached. Quite a bit better, eh? (Note: I truncated this one early and didn't let the whole download go through, since it seemed to have stabilized at 22 Mbps or so).