Bufferbloat: Dark Buffers in the Internet

Google Tech Talk (more info below)
April 26, 3011

Presented by Jim Gettys.

ABSTRACT

VOIP and teleconferencing often perform much more poorly on today's Internet than the Internet of a decade ago, despite great gains in bandwidth. Lots of fiber, cheap memory, smart hardware, variability of wireless thoughput, changes in web browser behaviour, changes in TCP implementations, and a focus on benchmarking Internet performance solely by bandwidth, and engineer's natural reluctance to drop packets have conspired to encourage papering over problems by adding buffers; each of which may introduce latency when filled.

Buffering mistakes have been made in all technologies: operating
systems, home routers both wired and wireless, broadband equipment, corporate networks, 3G networks and parts of the core Internet itself. The mistaken quest to never drop packets has destroyed interactivity under load, and often results in actual higher packet loss, as TCP's congestion avoidance algorithms have been defeated by these buffers. The lessons of the "RED manifesto" of 1997 have been forgotten or never learned by a new generation of engineers.

Full solutions require careful queue management, and that management should be everywhere; we no longer have the luxury to think that this is a problem solely of Internet routers. I will describe some of the mitigations and solutions to this problem, and how you can at least make your home network and systems behave the way they should.

More info at www.bufferbloat.net
Slides available at http://mirrors.bufferbloat.net/Talks/...

Speaker Info: Jim Gettys, Bell Labs
Jim Gettys is well-known as one of the original developers of the X Window System, and has long been active in open source and internet standards. His recent experiences with immersive telepresence applications exposed systemic implementation errors
in many Internet buffer and queue designs. He describes the journey of discovery in this talk.

Blog at http://gettys.wordpress.com/

• Category

  Science & Technology

[wwwhatsup]

I knew Google was getting ahead of itself but April 26, 3011? Wow!

[Hugh Jackson]

Unrelated, but your resume reads like the conquests of a computing demi-god.  Good work! :D

[fourbarposer]

Most (of the popular brand) LAN NIC drivers in Windows have Transmit Buffer Number or similar setting. On the low end on Realtek I can set it to 1, Atheros to 16.

However on USB wireless sticks, the settings are usually only related to the link/radio. It's however usually possible to download the drivers before purchase and see in the .INF file if it exposes the buffer# setting. Microsoft could also be interested in adding some note in their Windows HW Certification papers.

[slovokia]

The audio problems with this video probably were caused by mixing the output of two microphones together producing a flanging effect. If the microphone gains are set to produce similar levels interference results producing comb filer effects due to relative delays of the sound signal between the microphones.

[gael225]

What if we could eliminate middle hardward (excepting hubs) by using a near-perfect broadcast MAC? More like a DOCSIS distributed network, but with the ability to provision asynchronous flows for 1 to 1, 1 to many or 1 to all. Besides, how is anything 802.x going to handle trillions of devices without a massive infrastructure buildout. More at: vator.tv/company/ether2-1

[Jim Gettys]

No, because dark buffers that happen to be in places that can never be saturated can't ever come out to bite us.

But any that are a path in a location that saturates can/will bite you.

Also, where we can, we should be enabling AQM (aka RED). While it will have real problems in the "last mile" or your home, in classic router locations more interior to the network, it can be very effective. But at least we now understand why it has not been enabled everywhere it should be.

[Jim Gettys]

No mystery here: even web browsing (if you hit pages with lots of embedded objects) can easily induce transient bufferbloat.

But anything which is trying to work on top of a bloated network then has real trouble with estimating jitter, and can get quite confused.

[AntiProtonBoy]

Ran Netalyzr on my connection and crashed my router... hahaha!

[militantmindset]

so did he suggest that we need to firmware update the whole internet?

[telewebservices]

i like googletechtalks!

[Jim Gettys]

Sorry, I don't know my way around Windows. You can see if the network driver has knobs for setting the amount of buffering.

But by ensuring the bottleneck link is in your broadband connection (by ensuring your actual wireless bandwidth is always greater than your broadband link: e.g. by using 802.11n with care), and then using bandwidth shaping (aka QOS in many routers) as described in my blog. This will cost some bandwidth, but get you good latency.

[RuddODragonFear]

Vint, you are a cool guy, but... hey, too much introduction. :-)