Buffer bloat mitigation with OpenBSD pf
For an introduction to buffer bloat read more here http://en.wikipedia.org/wiki/Bufferbloat . My home network utilizes OpenBSD and the built in packet filter (pf). I use cable for broadband internet and found that if I tried to upload a large file my internet connecting became very unable with high amounts of latency. After utilizing tools such as http://netalyzr.icsi.berkeley.edu/blog/ it became obvious that I was suffering from buffer bloat. After doing some searching I came across using altq support in pf to try some configuration changes to reduce the buffer bloat in my configuration. I added a queue for my external network card with the following:
altq on $ext_if bandwidth 1Mb hfsc queue { bb }
queue bb bandwidth 100% qlimit 9 hfsc ( default )
and the corresponding rule to tag outgoing traffic for that interface to this queue
pass out on $ext_if keep state queue( bb )
My home connection is advertised as 5Mb down, 1Mb up. In testing I get about 1.1Mb up so I setup my outgoing queue to limit outgoing to 1Mb. A typical setting would be 97% of maximum. One of the most important values in the queue setup is the number of buffers. Mine is currently set at 9. This is how I determined what the value should be.
Maximum upstream bandwidth in packets is upstream bandwidth in bytes / size of a packet. In my case 1000000/8 = bandwidth in bytes / 1460 (size of packet) which yields 85 packets a second. So if I set my buffer size to 85 I should have about 1 second of latency. In my case I like my latency low so I divided by 10 to try and get a 100ms latency under full upstream use which is 8.5, which I rounded up to 9.
So how does it work?
Average pings to slashdot.org in milliseconds
Idle connection: 23.8 ( 0% packet loss)
Maxed upstream use: 95.1 ( 0% packet loss)
Maxed upstream without altq: 2083.4 (10% packet loss)
Quite the improvement! Thanks to the information on https://calomel.org/pf_hfsc.html for helpful tips.