Speed is hard [pdf] | Hacker News Reader

Speed is hard [pdf]

mirror.csclub.uwaterloo.ca

29 points2510c39011c515 comments

> Wireless is way too slow

Way too slow for gigabit internet, not way too slow for what most people actually need.

T-Mobile LTE in one city had an average of 28 Mb/s, enough speed for Netflix 4k, and 95 Mb/s max observed speed. I hate to support cable monopoly arguments in any way, but really for most people 100 Mb/s is overkill.

Of course the cell network couldn't support everybody's internet use with the towers covering such a wide area as today, but wouldn't fiber to the neighborhood and then wireless within it make a lot more sense than running expensive cable to every individual house? It wouldn't be gigabit, but if I could use uncapped tethering (due to smaller tower coverage area) that would be fine with me. T-Mobile LTE is plenty fast enough for anything I use the internet for.

Lets assume you get 100 ms higher latency with wireless. And you make around 100 clicks per day. Over your lifetime, you would have lost one hour of your life because you used wireless! When you are on the death bed you will think about all the things you could have used that hour for instead!

Lets take it a step further and count everyone who is using wireless today, that would be millions of hours lost. Think about what those millions of hours could have been used for instead!

That hour might be nothing for you, but for the whole humanity, it's a lot!

Sure, but reading these slides already used up a good chunk of the hour applying its lessons would have saved me...

The 0.1s wait time is less important than interaction error that happen due to latency in a dynamic UI. Also, surely some network round trips are not due directly to user input, but are still important to coordinate with other events.

100ms of network latency wastes a lot more than 100ms of user time per click, because loading a web page involves multiple round trips to the server. In the case of HN's front page, it took about five.

That's an interesting perspective. I mean, today, it seems impossible to imagine any use case for > 100Mb/s. After all, it can stream 4k HD video!

But I think the very way we use computers is restricted by what is possible on commodity infrastructure. The fact that 100Mb/s is overkill for many customers is simply because the average speed is much lower, so nobody has bothered to develop applications that make use of that additional bandwidth. If 100Mb/s became something everyone had, I'm sure we would start to see applications that require such speeds to be useable. I don't know what those applications would be, but I'm sure they would come to exist.

A 56k modem is overkill if you're browsing a text-only internet, but it would be pretty hard to use the internet of today with one :).

In the 56k days people could imagine sending audio one day, and dream about video. If we can't imagine a compelling use requiring gigabit, maybe there just isn't one.

I'm hoping that some experts here could clue me in, because with wired U.S. average according to Akami at 11.5 Mb/s it seems that a neighborhood wireless would be plenty fast for at least the near future, and much cheaper to build and maintain. Having fiber direct to the house would be better of course, but to me it just doesn't seem worth the cost.

Maybe I am doing the math wrong, but I think the plot "Minor Violations of the Laws of Physics" has the Shannon Limit 2x smaller that what it should be. According to [1], for each 2x increase in SNR, the limit should increase in a full Bandwith. However, when we compare the plotted values at 45db and 60db, we see that SNR increased 15db = 5 * 3db => 2^5. That way, the limit should increase in 5 bandwidths. But the plot increased from 300Mbps to 400Mbps => 100Mbps. Hence the plot was made with a bandwith of 100/5 = 20Mhz, instead of the 40Mhz mentioned.

[1] http://en.wikipedia.org/wiki/Shannon%E2%80%93Hartley_theorem...

I'll be honest, I find this math very confusing (as I am self taught using flakey internet resources similar to the one we're discussing :)). But my understanding is that there is a 2x factor somewhere that affects Shannon limit calculations. I think it might relate to the "negative frequency" mirrored half of the fourier transform at baseband. Or maybe I'm just doing the math wrong.

In any case, the underlying concept is correct: a 2x2 or 3x3 MIMO radio can exceed the Shannon limit calculated for a single channel.

Here's an article about "speed matters" for those that don't think latency is important.

http://blog.codinghorror.com/speed-still-matters/

Does QUIC help with overcoming the bandwidth delay product? Are web page loads with QUIC enabled faster on 1Gbps vs 10Mbps?

The bandwidth-delay product is a limitation of physics so you can't exactly "overcome" it. But because QUIC is implemented in userspace instead of the kernel, assumptions the kernel makes (like "the bandwidth-delay product will never be >= x") can be avoided in a cross-platform way, allowing for interesting, low-cost experiments that can later feed into improved TCP implementations (which take much longer to deploy).

Great read, thank you.

Speed is hard, but low-latency is even harder.

Did you read the pdf? It does talks about latency.

Reply on news.ycombinator.com