|Oh well, at least it's
Everything here is
my personal opinion.
I do not speak
for my employer.
Happy New Year
The first testimonial for djb redo
...other than from me, that is.
By comparison the "build-scripts/util" file in my "redo" branch is about 60 lines long, half of which is the non-Qt-specific "compile" shell function. Granted, my solution would be a lot bigger if it was as portable and comprehensive as the SCons plugin, but at least this approach was feasible for me to create in an afternoon.
Take that, scons.
(More about my redo project)
Notes on "The Yishan Tenets of Engineering Management"
Yishan Wong from Facebook posted about some of his most important ideas on Engineering Management.
I largely agree with his ideas, although there were a few things I obviously haven't thought enough about. I selfishly enjoy that kind of article, because it makes me feel simultaneously like I'm smart and like I'm learning something.
For example, on hiring managers:
I found that quote really striking because I have literally heard and accepted (and to my regret, even repeated) the "preparing the organization to scale" claim many times. Anything justified by this claim really has always been a cause of failure rather than success, but I hadn't noticed the correlation until he pointed it out.
(Ironically, I've reached the point in my career where various people who want to hire me think I should parachute in to be a manager for their engineering groups. Ouch.)
Relatedly, he has some specific comments on "process" as used in growing companies:
I have always hoped that the above is true, but never been able to prove it. I've certainly noticed a correlation between "high process" and "slow," but correlation doesn't imply causality, and slower isn't always worse... right? At least this provides one more vote for my preferred viewpoint, which is that if you have the right kind of culture, your company doesn't need as much "process" as you think you do.
Of course, I don't know much about the inner workings of the Facebook team; I don't even *use* Facebook. (Yeah, I know, I'm a luddite.) Based on my prejudices, running a company "that works like Facebook" - for example, in terms of revenue model and ethics - is not very high on my priority list. But oh well, I'll take wisdom where it comes.
Though in fairness, I now have to note that this creates (or maybe strengthens) a correlation between wisdom I like and companies I don't.
Finally, here's a quote that really got me thinking:
I have observed that different people have different levels of tolerance for chaos. But I'd never considered that even when someone doesn't *like* chaos, their company might be more productive by taking on the chaos anyhow.
That idea isn't obvious, but think of the consequences! A group of excellent people successfully doing what they're good at, in a well-oiled smooth-running company that doesn't make mistakes, can still be outcompeted by a bunch of goofballs who are literally screwing up a hundred times a day. All because the latter people can work with chaos - even when they don't like it.
Yishan's whole article)
bufferbloat (and a Net Neutrality rant in the footnotes)
Jim Gettys writes about "bufferbloat", the tendency of operating systems to keep increasing network buffer sizes so that now, with Internet links faster than ever, browsing often seems slower than ever.
His series of articles is a great introduction to the concept that I've been ranting to people about for years. Usually their eyes glaze over. What can you do? But his article is a bit more beginner-level, so maybe you'll have some idea what he's talking about.
This article, on the other hand, you probably won't.
Back in September, I was "lucky" enough to live for a month in someone else's apartment - and that someone was a cheapskate, so they subscribed to the cheapest possible Telus DSL plan, 256kbits/sec (symmetrical, for a change!).
On that connection, which is at least 4x as fast as single-channel ISDN or my old 56k modem (on which I used to multitask like crazy), it was totally impossible to do two things at a time. Loading a single web page totally killed interactive performance, to the point where we would have 8+ second ping times if someone had Google Reader open in a web browser tab. That same Google Reader that killed everything didn't even *work*, because it opened multiple sessions at once, and I think something in there somewhere must have had a 5-second timeout. Hilarity ensued.
Obviously, I concluded, our ISP has hard coded their DSL buffering based on the idea that we have a fast modem, even though we don't. At 6 megabits (about 23 times faster), that maximum 8-second delay would have been only 348 milliseconds... a mostly acceptable latency.
(Even if you have a fast DSL link, you still have the problem; most ISPs configure their DSL connections with symmetrical buffering, ie. the same sized buffer at each end. But even a 6 megabit downstream DSL link usually has a much lower upstream bandwidth - a megabit if you're lucky, or much less if you're not - so the same buffer size means a much higher delay on upload than on download. That's why on a normal link, *uploading* stuff kills performance much faster than *downloading* stuff. It's just misconfigured buffers, pure and simple.)
The bad news is that by the time you get into Jim's third article article in this series, he makes analyzing it all seem like rocket science. Bufferbloat is not rocket science at all. It's one single misconfigured number on virtually every DSL router (and now most cable modems too; technology apparently gets worse over time) everywhere. And unfortunately, it's often at both ends, yours and the ISP. And even when it's your end, you often can't reconfigure your DSL modem because it's locked down.
I heard a rumour once that ISP's do this on purpose: using a bigger buffer lets you squeeze a couple percentage points more bandwidth in stupid reviewers' benchmarks, and reviewers "of course" only ever download one thing at a time. You can evaluate bandwidth objectively, but lagginess during a download is mostly subjective. So perversely, an ISP that lowered their buffer size to make their connections suck less would lose out in third-party comparisons. Sigh. Having had my own products reviewed by morons in the past, I can't even fault them if that's the case.
On the other hand, the ISP here had no incentive to tune our cheapest-you-can-go 256kbit DSL link. In fact, they had an incentive to make it suck as much as possible, so we'd stop being such incredible cheapskates and actually buy one of their bigger, profitable plans.
So anyway, with all that suck out of the way, I have a message for the world:
You can fix it, and you don't need your ISP to not screw it up.
The bad news is that what you do need is a miracle :) Just kidding. Sort of. You don't technically need a miracle, but you need to figure out Linux's tc (traffic control) tool, which is about as close as you can get.
Let me jump ahead and say that, with a lot of fiddling back in September - let's just say I spent *most* of my September learning about this - I managed to get my 256 kbit connection to upload/download 10 files simultaneously while having a flawless Skype call, and my ping latency to Google never exceeded 250 ms. Were the downloads fast? Heck no, 256 kbits still sucks. But they were each about (my_bandwidth-skype)/10, which is to say, as good as it gets.
In case you don't already know, what you need is to insert another router (if you don't already have one) in between your home network and your DSL modem, and enable something called traffic shaping. On a typical DSL modem, you mostly only need to care about upstream traffic shaping, since that's the one that normally kills your performance (especially with BitTorrent); if you're picky, like me, you might also play with downstream shaping (called "policing" since it works totally differently), though the Linux code for that is not nearly as complete as the upstream stuff.
Here are some things I learned:
First of all, about 90% of traffic shaping setups that people try are absolute disasters. They make things much, much worse than nothing at all. It's a bit of a black art. It actually *shouldn't* be a black art, but there's something a bit funny about the whole department; everybody doing it seems to have taken user interface lessons from Cisco. The Linux 'tc' tools (kernel and userspace) were written by some kind of super genius in Russia, which is great as far as performance goes, but absolute death in terms of usability and documentation. I mean, the help message from the thing is literally a dump of its BNF grammar. If you're a Russian super genius, that's probably all you need. If you're me, you cry.
Given that, the terrible state of the documentation - all written by other people, Russian super geniuses don't need no stinking documentation - is actually kind of forgivable. Except in the cases where they obviously haven't even tried their own advice. Oh, sure, the sample scripts all *run* without errors, but they just plain don't do the things they claim to do. I'll get to that in a minute.
So, ok. Super complex UI, lots of fiddly magic numbers, and no documentation. You're sort of doomed. I know, I'm sorry. That's why this article is more of a rant than a howto. I won't keep you in suspense: you're only halfway through this article, but it doesn't get any better. Feel free to stop now before you get as depressed as I am.
BTW, all the "user friendly traffic shaping" scripts I could find make horrible mistakes in configuration and do lots of weird things. If anyone remembers the Nitix "Active Queue Management" feature and that it would do insane things sometimes... now I know why. Because we didn't spend enough time figuring out what the heck those scripts were doing, and realizing that they're doing it wrong.
Actual experimental observations
With upstream shaping, you can set it to about 99% of your upstream bandwidth and life is good. This makes sense, of course, because some traffic can jump straight to the head of the queue, so you don't really have latency concerns. (By the way, with an idealized perfect queue, there's no reason to ever limit the queue size. But such a queue doesn't exist.)
With downstream policing, you have to cut bandwidth usage to 80% or less, or the jitter will kill your latency "sometimes." (This might not matter as much on faster-than-256k links. At 256k, even three full-sized packets enqueued at once can cause 140ms of momentary delay. If you have ten downloads at a time, the *average* case is about 10 packets at once, or 458ms, and it gets worse from there.) So this 80% number isn't really fixed; it varies based on how many connections you have going at once.
If you're massively exceeding your allotted bandwidth (that is, always, if you have a super-slow 256k link like mine, or if you have a normal link and you're doing a lot of torrenting/etc) then enabling ECN on all your client machines, and using an active queue like RED on your router, can make a huge difference versus no ECN.
ECN's biggest effect is in improving interactive traffic when there are high-bandwidth connections around (interactive TCP responds *very very badly* to packet drops, since it doesn't have any other packets outstanding, so it doesn't get the instant re-ACKs caused by later packets... long story, but losing a couple of packets can cause multi-second delays on interactive traffic, and only millisecond delays on busy sessions). The good news is that this means you can enable ECN on *just* your own client that you use for interactive traffic, and your results are improved, even if you don't want to ECN-enable every single computer on your network. (This isn't cheating; you aren't hurting anybody else's performance by turning on ECN for only one machine. Like magic, ECN is just always an improvement, and always fair. Nowadays, chances are that the reasons ECN is off by default don't apply to you anymore; try turning it on and feel the love.)
If you do traffic shaping/policing properly, there is little need to actually *prioritize* traffic; just letting all the streams share fairly is usually enough. I was able to have a Skype call just fine on my 256k link with 5 downloads going at once with traffic speed limiting but no prioritization. With 10 downloads, 1/11th of my bandwidth was just too small to get good sound quality, so prioritization was the only way. But come on, if you have a 10 megabit link, do you *really* think that's your problem? Plus, you can't do prioritization of your downstream - that's something only the ISP can do, and they won't, for very good reasons.1 So you're lucky it turns out not to matter.
So the rule of thumb? Get your basic queuing right *first*, and worry about prioritization *second*. A lot of people get those two mixed up. Prioritization is worthless when your queue is doing stupid things.
Next, beyond simple idiotic oversized DSL buffers, the *real* problem with BitTorrent is that it uses so many simultaneous TCP connections. If you have 50 of them at once going at full speed, everything else will get 1/50 of the bandwidth at most. As far as I know, nobody has ever found a really good trick for protecting against this. (You could group traffic by IP instead of by IP:port, but that would penalize anybody behind a NAT, so it's no good.)
Linux's traffic policing (ingress) layer can't do ECN. This is a terrible, terrible oversight. Someday I may try to fix it, if someone doesn't beat me to it.
Some people suggest shaping the *outgoing* queue on your *local* interface as an alternative to policing the *incoming* queue on the *remote* interface. Same thing, right? Packets still get dropped when they exceed the bandwidth, but now you can use all the fancy queue types, including RED, which allows you to use ECN for incoming traffic! Except this setup results in total crap; I wonder sometimes if the people writing these howtos even try their own advice. When receiving packets from a slow link, you've already paid the latency costs; if you now create a queue before sending the packets onward locally, you're just adding a bunch more latency. And the quality of (for example) your RED tagging varies proportionately with the added latency; if you keep the useless queue short, the policing quality drops because the algorithms aren't made for that. A much better solution would be a policing ingress queue that would "virtually" drop packets using ECN when possible.
I think you could make a "virtual RED queue" that doesn't actually delay or even store packets - RED isn't about delaying packets anyhow, it's about dropping them before the queue is full, and in fact it's better to ECN tag them instead of dropping them, so why not virtually drop them when your virtual queue is virtually full? It would be about as good as really dropping them from a real queue that's really full, but with no added latency and no actual packet loss. Someone could probably write a Ph.D. thesis about this, which I'm guessing is also why nobody has implemented it.
Note that regardless of the above, an *outgoing* queue used for traffic shaping does *not* increase latency. It always releases the packets (just slightly less than) as fast as they can go. So even though the queue usually has a bunch of stuff in it, there would have been a queue with stuff waiting in it anyhow; traffic shaping just makes it smarter. This fundamental difference is why using your outgoing queue algorithms for ingress traffic totally doesn't work.
TBF vs. SFQ
Whatever you do, don't try to combine TBF (token bucket filter) with SFQ (stochastic fair queuing), because the end result is dramatically bad. Unfortunately, some of the Linux howtos actually show you how to set this up as one of their examples! But it's fatally flawed.
A basic non-prioritizing TBF will simply drop the next packet if there's been too much data lately; statistically speaking, the busiest connections have the most packets, so they're the most likely ones to have a packet dropped. Interactive sessions, with just one packet in a blue moon, have a very low probability of dropping a packet. Pretty good, right?
That's TBF by itself. I mean, it's not perfect: even my 1/100 speed interactive session will get screwed for 1/100 of the packet drops, even if my busy connection is 99/100 packets. What we really want is for *all* the packet drops to punish the high-bandwidth guy. Still, it's pretty good considering how simple it is, and being wrong 1% of the time isn't so bad.
Now, let's look at SFQ, which is a very clever invention. SFQ works by sorting each session into one of, say, 100 "bins." Then it feeds the first packet in each bin to the output queue, looping through the bins in sequence. When a new packet arrives, it sorts it into a bin, and drops the packet only if *that* bin is full, but other packets, in other bins, don't get dropped. The trick is that packets are sorted into bins based on the hash of their (srcip,srcport,dstip,dstport) tuple, so on average, each connection gets its own bin. Assuming there are a lot more bins than connections, the end result is that you *never* sacrifice your 1/100 interactive session by accident; the only packets that ever drop are the 99/100 busy session, exactly the guy you want to slow down. Sweet! And it really works, too... if your network device never drops packets from its own queue, which is a safe assumption for physical devices.
Now let's hook SFQ to a TBF instead of a physical device. SFQ cleans things up, sorting the packets in the order A,B,A,B,A,B (where A is your interactive session, and B is your high-bandwidth session). It feeds them into TBF, which queues, say, 10 packets at a time. When the tokens run out, it drops a random one of the 10 packets, and lets in a new one from the SFQ. Which one will it choose randomly? Well, given that SFQ has inserted a whole bunch of fairness into the queue and the packets are no longer bursty... your interactive session has about a 50% chance of having its packet dropped by TBF.
OH MY GOD IT'S SO TERRIBLE.
And I'm not just making this up. I tried it. I really did. I gathered statistics. I analyzed packet dumps. I learned what all the crazy magic numbers did (and tc has a *lot* of crazy magic numbers). And the conclusion was: it works just like I said just now. And it sure does *completely* tank your interactive performance. TBF by itself is actually okay (1% error), and SFQ is okay (~0% error) if that's your problem, but TBF+SFQ is never, never okay (50% error).
Two rights make a wrong? That's a new one.
Now, you could probably fix this by writing a different sort of TBF: one that accepts packets into its queue on a particular schedule, but assumes the guy feeding it packets will be the one doing the dropping; my alternate-universe-TBF would never drop packets itself. This parallels my "virtual RED queue" suggestion up above, but beware that SFQ+RED is probably just as bad as SFQ+TBF.
By the way, just putting things in the opposite order - TBF then SFQ - doesn't work either. The SFQ always drains instantly, and without packets in the queue, it can't add any fairness.
Oh, and also, if you think about connecting PRIO (basic prioritization, ie. VoIP vs. bulk) to TBF... you get exactly the same problem. TBF will end up penalizing high-priority packets even more than it would without PRIO! Which is why some people find that enabling packet prioritization makes their Internet performance worse.
No, HTB is not the answer!
And before you say something, yes, I've read all the notes that HTB (hierarchical token bucket) is the newfangled big thing that will solve all your problems.
No it won't.
HTB is probably better than CBQ, the thing it was designed to replace, but neither of them actually solves any of the problems I'm writing about here. As soon as people start talking about traffic shaping, they always want to start talking about restricting one stream to 1 megabit, while the other one gets 5 megabits, but then we allow short bursts of xyz for not longer than abc every pqr, and then we subdivide it like this...
It's all wankery. You're sitting at home, your Internet sucks. You do not need to do any of those things. Moreover, even if you had the problems those people are talking about, you would still have the latency problems *I'm* talking about, and HTB gets you 0% closer to solving them. HTB is good if you need it. But chances are, you absolutely don't.
The real problem with traffic shaping...
...is that it's just too fiddly. The truth of the matter is that it's really just a few little numbers - really simple numbers, integers, like the maximum number of packets in a queue - that make a huge difference. Coming up with this small set of numbers requires a surprising amount of math, and if you don't get your math exactly right, you don't just get substandard results - you get horrifically bad results. A mis-tuned RED queue, as I learned in my tests, is both very easy to produce, and very destructive to use. Combining TBF, which is great, with SFQ, which is great, results in disaster. It's really mathematically hard.
The funny thing is that, if you could just reliably get the math right, the actual code behind traffic shaping is delightfully simple. Any tiny little routing box with a super crappy CPU and almost no memory can do it, and the results would be great.
It's one of those really frustrating computer science problems: most programmers just like to get down to coding stuff. If you work a little harder, your solution gets a little better. The world of traffic shaping isn't like that; it's not a lot of work, but the outputs also aren't proportional to the inputs. It's a step function from crappy to awesome, with nothing in between to give you a hint that you're on the right track.
Footnote: net neutrality
1 I said up above that ISP's won't do packet prioritization at their end of the link, for "good reason." That reason comes down to the whole famous "net neutrality" debate; should ISPs prioritize traffic or is every stream equal to every other stream? You might have heard of the net neutrality debate, but every time I hear it, I always think: wait, which side of this debate am I supposed to be on, anyway? I can never tell.
So far my answer is, it all sucks. In short, everyone can agree that their phone calls should take precedence over their bittorrent downloads. They can probably even agree that their phone calls should take precedence over their web browsing. And you know, most people can even agree that *other* people's phone calls and web browsing should take precedence over their own bittorrent downloads. Most of it is easy; most of the people spamming you about net neutrality only tell you about the easy bits.
But the thing nobody can agree on is... what *is* a phone call? And what *is* web browsing? And how can your ISP tell the difference so they can prioritize it the way we all agree is the right way?
Packet prioritization on your local network is no big deal. (It's useless, but easy; by the way, if your wireless router has a "wireless QoS" option, just turn it off. I guarantee you don't need it, and it definitely doesn't do what you think it does.)
Conversely, packet prioritization on the open internet is a giant security hole. If you can tag a skype session as "high priority", then what's to stop people from tagging their browser sessions as high priority? What's to stop some annoying hacker from blasting a million "high priority" packets at your IP address, completely filling your queue, which prevents any low priority packets from getting through *at all*, thus destroying your ability to access email or the web? Nothing can prevent this but your ISP. And so your ISP is stuck: do they honour those packet priority tags, or not? If they do, giant security hole. If they don't, skype calls are a bit less clear. On the open Internet, they pick the latter because the former is simply not acceptable to anybody.
Now, at least in Canada, all the ISPs are *also* in the VoIP business nowadays. They offer special boxes that do phone calls over their Internet link, in parallel with your cable/DSL modem that does Internet stuff. And they prioritize *that* traffic, right? Which is totally unfair to Skype, right? Totally unfair? Yes, absolutely!
But unfortunately, also justifiable for technical reasons. I'm not talking about business here; that's another question. But the idea is that you *obviously want* your phone connection to take precedence over the Internet connection; your ISP controls their network, the cable modem, and the phone box; and so the ISP can safely tag the packets - in a way your cable modem can be configured not to do - so that you and anybody else can't fake high-priority packets and cause trouble.
In short, barring insanely complicated inventions (ie. intserv, "integrated services", the failed opposite of diffserv) you or anyone can only securely implement prioritization on a network you completely control.
Unfortunately this gets the net neutrality police up in arms, because now BigCo's VoIP service has more reliable performance than Skype. And oh boy, does it ever; the sound quality is pretty much perfect, all the time. And yeah, that helps BigCo reinforce their monopoly position, and that totally sucks for Skype and competition and consumers. But make no mistake; if you pass a law requiring Net Neutrality, then you'll just make the VoIP service suck; you won't make Skype better. Or if you pass a law requiring ISPs to implement diffserv, then the whole Internet will collapse because a few idiots will abuse it by making their torrents look like phone calls.
And how about those special deals for Youtube and Netflix? Are those fair? Nope. But because those companies have special arrangements with the ISPs, the ISPs *can* prioritize those live video streams over random Internet crap without creating a giant gaping security hole. And let's face it, average consumers *do* want their Youtube and Netflix to be prioritized over their random web browsing, because glitchy video sucks, and glitchy web browsing is mostly unnoticeable. But is it fair to J. Random Bob's video sharing site? Of course not. Is there even any way, technically, for ISPs to offer the same service to J. Random Bob? No, and that's because the universe is fundamentally unfair.
With net neutrality, your choice is between suck and unfairness. Both options are terrible. And that's why neither side has won the debate and governments can't figure out what to do.
Oddly, it doesn't even matter that ISPs stand to make a lot of money by auctioning off priority to the highest bidder. We could deal with that; that would just be plain simple corruption. Take a bribe, or throw someone in jail, or whatever, but it would be over with by now.
There is some good news, though: I read somewhere (I can't find the link right now) that in Internet backbone-scale tests, simply overprovisioning bandwidth by around 1.6x (ie. 60% more bandwidth than you need) reduces jitter and latency just as much as packet prioritization. So if available bandwidth keeps going up at the rate it has been, unless some accursed person invents something even more bandwidth-wasteful than video (what could possibly do that?!) then this whole debate should end itself in a few years. When in doubt, use brute force.
Of course, even that does you no good if your router's queue is too big.
bufferbloat vs. wireless networks, and other stories
It seems my previous article on bufferbloat struck a chord with some people. Most people, quite rightly, didn't know what to do with it and thought it was super confusing. (It was. It was rambly and poorly edited, and reading the responses mostly reminded me how disorganized all my thoughts on the topic were. Oops. It turns out disappearing into a hole and studying something for a month straight can quickly make you lose perspective on what the average person does/doesn't know about TCP/IP and traffic shaping. Sorry about that.)
On the other hand, some people actually think there's useful information hidden in there, which there is. Other people thought I was just wrong, which I'm not; that article, confusing as it was, isn't *conjecture*, it's a disjointed disorganized overly-summarized report of my actual careful experiments and observations.
Now, vague inexact reports of my careful observations are just as useful to you as totally wrong conjectures, so I apologize for that. Perhaps I'll try again later.
Meanwhile, I do want to correct a couple of misconceptions in particular, just in case people *did* believe me, because there are some things I *wasn't* claiming because they aren't true and/or I don't know if they're true.
Wireless networks have problems, but they are not these problems
My article was specifically about a very simple network topology that looks like this:
computers -> wires -> router -> DSL/cable/etc -> internet.
I didn't talk about wireless anywhere in the article, somewhat on purpose, because wireless has its own problems and those new problems confound your ability to do straightforward experiments to see the original router problems I was talking about.
Now, it's safe to say that if your wireless speed is much much more than your router speed, and all your wireless traffic is internet traffic (ie. not talking to other wireless devices on your LAN), then everything in my article remains true. Any wireless problems are dwarfed by problems caused by your router's crappy queue, and can be solved by doing traffic shaping correctly.
TCP is smart, so it knows not to send data any faster than your router can handle. Since that's much slower than your local wireless network, the only queue that ever fills up is the one at the bottleneck: at your router or the ISP's router on the other end. All the boxes on your WLAN will have essentially empty queues, so it doesn't matter how their maximum queue length is configured.
However, let's imagine that this isn't the case on your network; nowadays you might have 20 megabit cable, or 50 megabit FIOS, or whatever, which means your Internet connection is actually faster than your wireless network. Or maybe your Internet connection *is* a long range shared wireless network, in which case your wireless speed is obviously not much much greater than your router's speed, as we had been assuming. Or maybe you're just connecting two computers together on your LAN, with no router involved at all. What happens then?
What happens then is, well, a whole different story than I had been talking about. And so of course the solution will have to be totally different.
Now, as it happens, I've been thinking about creating a new startup to build the ultimate wireless router - one with built in traffic shaping *that works out of the box*. So while I don't have as much concrete test data for wireless as I do for wired routers, I do have quite a bit of information that comes from studying the situation in theory and a bit of practice.
Here are some things you need to know:
First, various operating systems and drivers have oversized network buffers on their wireless card drivers (and on their wired ethernet drivers too, in some cases). Since your internet router is no longer the bottleneck, suddenly the *client machine's* buffers are what matter, and those buffers are oversized. So yeah, if you start a big transfer, your interactive performance will suck. Solution: reduce the buffer size. Or install traffic shaping on your client machine, but that's a bit silly since it's your own LAN, so you should just fix the buffer size.
If you're downloading from a really dumb local server that has oversized transmit queues, same problem. You should get the server admin to lower the queue size. I guess your admin might be mean and refuse to do that, in which case you can improve your personal interactive performance when communicating with that server by installing traffic *policing* (downstream) on your client, but that's gross and fiddly.
Incidentally, traffic shaping/policing on a wireless network is totally disgusting because the available bandwidth is *variable*. How do you set the rates to 99%/80% of the bandwidth when the bandwidth keeps changing? You can't, at least not without having a direct line to the network driver to find out the current bitrate. And anyway, it's all stupid, because the right thing to do is simply lower the driver's tx queue size. Do that.
Next, rumour has it that if one machine starts blasting data all over your wireless network at top speed, this will hurt performance of other machines on the same WLAN. I personally have never experienced this, so warning: I'm now talking theory, not experimental evidence. But if it were to happen, I claim that this can't *possibly* have anything to do with the queue size of any of your devices. Why? Because no matter how big its queue, the wireless device on any computer can't send faster than its maximum transmit rate. A shorter or longer queue affects latency *on the machine with the queue*, and *on any machine forwarding through that machine*, but to everyone else, it just looks like the same packets coming out at the same speed.
(If your problem is that one machine on your wireless network starts blasting tons of data and overwhelms your router, well, yeah. Your router queue is too big. But if it overwhelms your router, we're in the case we talked about last time, not the current case where your router is too fast to ever be overwhelmed.)
Now, just because the problem isn't caused by your queue length doesn't mean I'm claiming it doesn't exist. There are lots of reasons the problem might exist, *especially* on a wireless network.
First of all, 802.11 wireless networks retransmit packets when they get dropped. This sounds obvious - doesn't everybody just retransmit stuff? - but no, it's *very strange*. In fact, ethernet *doesn't* retransmit stuff. If it sends a packet and the packet is lost due to noise or because the receiver has no buffer space, then tough. The packet is lost. It's TCP's job to retransmit it.
802.11 wireless networks don't work the same way. The reason is this: TCP *depends* on information about how many packets are lost and when. TCP assumes it's on a wired network, which is generally reliable, so when a packet is dropped, it's almost always because the receiver (or actually, any router along the path) has no more buffers. If that's true, it means the router is overwhelmed, so TCP needs to slow down... which it does. And that's how the whole internet works! That's why you don't need to know the speed of *my* DSL modem in order to send data to *my* DSL modem at exactly the speed my DSL modem can tolerate; no more, and no less. It's very cool.
Anyway, wireless networks are way less reliable than wired ones, because wireless networks are much more susceptible to random noise. So packets get dropped all the time, and it's not because you're sending too fast, it's because you're unlucky. Thus, the 802.11 standard calls for devices to retransmit packets that were lost at the physical layer. They can still be dropped by the receiving device, but only *after* they've been acknowledged by 802.11, which means the transmitter knows you got them. So in 802.11, there are two kinds of packet loss: the kind you're supposed to fix (noise), and the kind you aren't (buffer full).
Generally it's the driver's job to do the 802.11 retransmits. This calls into question: how many times do you retry transmitting? How long between retransmits? What if the noise was caused by someone else's retransmits, and we get into lockstep, constantly ruining each other's packets? On the receiving end, how and when do we report which kind of receive failure is which?
Answer: it's complicated. But it's obviously easy to do wrong. I heard a rumour that some Linux drivers will retry the same packet up to 256 times in a row if it doesn't work; that seems rather desperate to me. I can also imagine bugs where a driver, finding the kernel's receive queue is full, reports the 802.11 packet as corrupted ("please retransmit") rather than as delivered-but-then-dropped. If it did that, it would utterly destroy TCP's behaviour and potentially clog the network really badly. Maybe this bug exists; I have no idea, I've only heard rumours. If it does, then it would surely be a problem. It would also have nothing to do with queue length, and could not be solved by any form of traffic shaping.
Which brings up another really important point: the whole Internet actually only works at all because of a truly amazing level of voluntary cooperation. Anybody anywhere could deliberately mis-tune their TCP layer and get faster downloads and uploads than anybody else, at the cost of messing up things for everybody else. Similarly, a WLAN driver that mis-reported dropped packets could get much higher thruput than anybody else on the WLAN, at the cost of screwing over everybody else.
TCP is kind of like the "prisoner's dilemma" - as long as we all cooperate, we all win, but if any of us defects, we might as well all defect, because the party is over. And as with experimental tests of the "iterated prisoner's dilemma" in real life (where you play the game with your partners over and over, rather than just once), we have mostly converged on not cheating. Because if we didn't, someone else would cheat too, and the party would be over.
Overly long transmit queues are not defection, by the way; they're just stupid. You don't get better performance at other people's expense by using overly long queues. You just give yourself worse performance at your own expense. Nice work, genius.
So that's not this. And to repeat myself, queue length is also not what would cause one wireless device to steal bandwidth from another wireless device. 802.11 retransmit problems? Yeah, that could cause problems all right.
Now, even if we assume our 802.11 drivers don't have bugs (other than queue length, which doesn't screw anybody but the perpetrator), which might even be the case, there are *still* some obscure reasons that one station's WLAN traffic might stomp on another's. Here are a couple that I know about; there may be more.
1) Ethernet-style collision detection (which is used, sort of, sometimes, in 802.11) is prone to timing problems; the more people you have on your network, and the faster the network, the more collisions you'll have. Background: in original 10MBit ethernet, wire lengths were limited so that the first byte of a packet A would reach all nodes by the time the last byte had been sent out. It was still possible that a second node would start transmitting its own packet B between the first node starting to send A and the second node start to receive A; however, the timing was arranged so that both A and B would know by the end that they had stomped on each other, so they could wait for a random delay and then retransmit.
Unfortunately, with 100MBit and faster networks, this collision detection got a lot worse, because the same packet was 1/10th the "length" (if you think of a packet travelling through a wire as a wave moving at the speed of light, the distance from start to end is the "length" of the packet), and that length was so short that it wasn't really reasonable to require nodes to be physically so close together. That's why hubs were sensible in the 10MBit days, but rapidly fell into disuse in the 100MBit days in favour of switches, which don't need collision detection because they just give every node its own wire.
With wireless, you have no choice: everybody is on the same "wire", ie. the airwaves. You can sort of help by using multiple channels, which 802.11 does, half-heartedly. (There are only 3 non-overlapping channels in 802.11's North American standard, at least, which means three packets at a time in an optimistically configured topology.) There are interesting mathematical improvements on this (eg. frequency hopping and ultrawideband transmission), but 802.11 doesn't use them; cell phone networks do, nowadays, because they have so darned many people using them at once that nothing else would suffice.
*Anyway*, with 802.11, you're back to old-fashioned collision detection, only now you have outside noise, uncontrolled distances, random signal reflection, *and* fast transmit rates. It's awful, and frankly, I'm amazed it works at all.
The collision detection algorithms have improved somewhat since the ethernet CSMACD (carrier-sense multiple access collision detection) days, but not much. And the various 802.11 standards - I haven't read them all yet - seem to define multiple different ways around the problem. One of them is actually, ha ha, a token-ring like scheme: the wireless access point assigns timeslots to each node, and the node is only allowed to transmit during its timeslot, thus avoiding the possibility for collisions (while also bringing back all the bad stuff about token ring that killed it in the first place). I'm a bit fuzzy on when this mode is enabled and when it's not; I think it's mostly not, so this is more academic than useful.
So anyway, collision detection is hard. And the more you transmit, the more collisions you'll have. The more different nodes you have transmitting, the more collisions you have. The more your driver's auto-retransmit algorithm is broken, the *vastly* more collisions you'll have. And so on. One busy station with a wrong retransmit algorithm could really ruin it for everyone. Of course, like I said, I have no evidence that such a thing exists; I've just heard rumours.
2) This one is even more insidious. It's called the "hidden node problem." It works like this: station A can talk to the access point, and station B can talk to the access point, but A and B are on opposite ends of the building, so they're too far away to see *each other*. So let's say they both decide to start blasting data at the access point at the same time at full speed.
No matter what collision detection method they try to use, they'll never know that every single one of their packets is colliding with the other guy! As far as the access point is concerned, it's seeing a nonsense combination of signals from A and B. As far as A and B are concerned, all is well,1 except for some reason the access point isn't acknowledging any of their packets, so they have to retransmit.
If you're running a long-range wireless network, your access point probably has a really great well-positioned antenna, and the other stations probably don't. That means you probably have the hidden node problem in a *big* way. That means tons of horrible unexplained packet loss, which means tons of retransmits. And, rumour has it, those retransmits might be compounded by buggy retransmission algorithms. As the operator of such a network, your life, in short, sucks big time.
I don't have any experience running long-range multi-user wireless networks, although I have fantasized about doing so, in what now appears to be a rather obsessive-compulsive level of detail. And the hidden node problem is where my fantasies turn into nightmares.
I don't know the solution to this one. I do know three things that *might* be potential solutions, but they'd need to be thoroughly tested:
a) That 802.11 timeslot mechanism I mentioned above, if it actually was turned on and supported. (I read about it in the spec, and got the distinct impression that it was a great idea that wouldn't do me any good because I couldn't use it. But I now don't remember why. I think it's part of the 802.11 QoS spec, and even if routers support that, none of the standard workstation drivers do.)
b) Some projects (frottle, WiCCP) that actually do timeslicing at the user level. At least frottle has some user reports that say it dramatically improved results. However, it requires *everyone* on your network to run special software; if anybody isn't, they'll probably stomp on all the packets from everybody else.
c) An slightly insane specialized application of traffic policing at the access point. This requires that everybody talks only to the access point directly, never to other nodes on the WLAN, which is a pretty safe assumption since the hidden node problem means they can't see each other anyhow. Basically, if you do things exactly right, you should be able to slow down everyone's traffic so they all get 1/n of the total bandwidth, and you should be able to reduce traffic burstiness so those packets mostly statistically end up not colliding. If it worked, it would be a totally amazing hack, because you wouldn't have to fix the software on any of the stations, and you wouldn't have to use explicit timeslots; the timeslots would be implicit based on your traffic policing algorithm.
And no, the transmit queue length has no bearing on this problem. In fact, somewhat entertainingly, activating policing on the access point's WLAN link could actually force the transmit queues to be shorter (since the TCP window size would be shorter since policing is dropping so many packets), thus solving that problem too.
Assorted other clarifications about my last article
While we're here, note that:
1 You can simulate the hidden node problem at home while doing
the dishes. Go stand by the kitchen sink and turn on the tap. Now talk at
a normal volume to someone across the room, and have them talk back to you
at a normal volume. To them, you'll be audible; what they hear is about a
1:1 ratio of tap noise to voice, which is good enough to understand. But to
you, they'll be inaudible; by the time their voice reaches you, it has
decayed a lot (ie. by the square of the distance between you and them), but
the water noise is full blast, totally overwhelming their voice.
redo surpasses sshuttle in github popularity
Today my redo project (based on djb's redo design) passed my sshuttle project in terms of number of github followers, despite being significantly newer. bup still has about twice as many followers as either of them, though.
...and yes, that's right. I'm competing with myself. I like it better than competing with other people, because it keeps me on my toes: I always lose, but it's never depressing.
Moore's law and iPad-sized "retina displays"
Normally I don't bother to engage in Apple product speculation, but in this case, I have some actual knowledge (from a brief foray into working at a semiconductor company, some time ago) that might be interesting to share, and Apple is a good excuse. Of course, that was a long time ago - before LCD monitors had any significant popularity - and I'm just a software guy, so I might totally misinterpret everything.
But if I were to take what I think I know and conjecture wildly without looking at any reference material, it would look like this:
LCD-style displays (at least in that family; that is, a two-dimensional array of transistors that light up or darken or whatever to form an image) are all built using something like semiconductor crystal growth processes.
Semiconductor processes more or less conform to Moore's Law, one statement of which is: the price per transistor approximately halves every 18 months. (It is often stated in terms of "performance," but we have already long passed the point where doubling the number of transistors automatically improves performance. And in a display, we care about the literal number of transistors anyway.)
Why does the price per transistor halve? I'm not really sure, but the short answer is "because we get better at growing perfect (imperfection-free) semiconductor crystals at a smaller size." So there are two variables: how small are the transistors? And how perfect are the transistors?
Roughly, if you can pack more transistors into a smaller space, the chance that you'll run into an imperfection in the crystal in the space you use will be less; the error rate in a crystal is about constant - and very low, but not quite zero.
Transistor sizes ("processes") come in steps. These are the sizes you hear about every time Intel builds a new fab: 0.04 micron, and so on. When a new process is introduced, it'll be buggy and produce a higher rate of imperfect transistors for a while; with practice, the manufacturing engineers get that error rate down, saving money because more of the chips on a particular wafer of silicon will be imperfection-free.
By the way, that's how it works: you buy these wafers of silicon, then you "print" circuits on it using one of many extremely crazy technologies, then you chop up the wafer and make chips out of it. Some of the chips will work (no imperfections), and some of them won't (at least one imperfection). Testing chips is a serious business, because it can sometimes be really hard to find the one transistor out of 25 million that is slightly wrong sometimes (at the wrong temperature) because of an imperfection. But they manage to do it. Very cool.
Also by the way, multicore processors are an interesting way to reduce the cost of imperfections. Normally, a single imperfection makes your chip garbage. If you print multiple cores on the same chip, it massively increases the area of the chip, making the probability of failure *much* higher. For example, a quad-core processor takes roughly 4x the area, so if p is the probability of failure for one core, the probability of overall success is (1-p)**4. So if one core has a 90% success rate (where I worked, that was considered very good :)), four cores have only a 65% success rate, and so on. Similarly, a single-core megacomplicated processor four times as big as the original would have only a 65% success rate. But if you're building a multi-core processor, when you find that *one* of the cores is damaged, you can disable *just* that one core, and the rest of the processor is still good! So as a manufacturer, you save *tons* of money versus just throwing it away.
Remember those three-core AMD processors that people found out they could hack into being four-core AMD processors? I wouldn't, if I were you.
Anyway, where was I? Oh, right.
So every now and then, Intel will build a new fab with a newer, smaller process, and start making their old chips in the new process, which results in a vastly lower imperfection rate, which means far fewer of the chips will be thrown away, which means they make much more money per chip, which means it's time to cut the price. Along with improved manufacturing processes etc, this happens at approximately a 2x ratio every 18 months, if Mr. Moore was correct. And he seems to have been, more or less, although some of us might wonder whether it's a self-fulfilling prophecy at this point.
Now here's the interesting part: LCD-like displays, being based on similar crystal-growing methods, follow similar trends!
A "higher DPI" - more dots in less space - corresponds to a new crystal growth process. Initially, there will be a certain number of imperfections per unit area; over time, those wonderful manufacturing engineers will get the error rate down lower.
Remember when LCD monitors used to come with "dead pixels?" The small print in your monitor's warranty - at the time, maybe not now - said that up to 1, or 2, or 5 dead pixels was "unavoidable" and not grounds for returning the product to the manufacturer. Why? Because it was just way too expensive to make it perfect. Maybe the probability of one dead pixel in a 17" screen was, say, 80%, but the probability of two dead pixels was only 50%. If you allow zero dead pixels, you have to throw away 80% of your units; if you allow one, you only throw away 50%; and so on. Accepting those dead pixels made a *huge* difference to profitability. LCD-like displays might never have been cheap enough if we didn't put up with the stupid dead pixels.
But those were the old days. Nowadays, thanks to greatly improved processes and lots of competition, we've gotten used to zero dead pixels. Yay capitalism. So let's assume that from now on, all displays will always have zero dead pixels.
The question now is: how long until I can get an iPad with double the resolution?
Well, now we apply Moore's Law.
If you watched the Android vs. iPhone debate over the last couple of years, you saw a series of Android phones with successively higher DPI in the same physical area, followed by the iPhone 4, which had a just slightly higher DPI than the most recent Android phone at the time, but which happened to be double the resolution of the original iPhone, because the iPhone hadn't been trying to keep up. (Now that the pixels are so tiny that they're invisible, are Android phones still coming out with even higher resolutions? I haven't heard.)
The so-called "retina display" wasn't a revolution, of course - it was just the next step in displays. (Holding off until you could get an integer 2x multiplier, however, was frickin' genius. Sometimes the best ideas sound obvious in retrospect.)
Each of those successively higher resolution screens was a "new process" in semiconductor speak. With each new process came a higher density of imperfections per unit area, which is why smaller screens were the first to hit these crazy-high densities. We can assume that the biggest screens at "retina" density as of the iPhone 4's release - 960x640 in June 2010 - was the state of the art at the time. We can also assume that the iPad was the biggest you could go with the best process for that size at the time, 1024x768 in March 2010.
Let's assume that the iPad wants to achieve a similar DPI before it hops to the next resolution. I think it's safe to assume they'll wait until they can do a 2x hop like they did with the iPhone. However, the exact DPI isn't such a good assumption; people tolerate lower DPI on a bigger screen. But since I don't have real numbers, let's just assume the same DPI as the iPhone 4, ie. the same "process."
We want to achieve 1024x768 times two = 2048x1536 = (2.13*960)x(2.40*640) = about 5x the physical area of the iPhone 4.
Moore's law says we get 2x every 18 months. So 5x is 2**2.32, ie. 2.32 doubling periods, or 3.48 years.
By that calculation, we can expect to see a double-resolution iPad 3.48 years from its original release date in March 2010, ie. Christmas 2013. (We can expect that Apple will have secret test units long before that, as they would have with the iPhone 4, but that doesn't change anything since they do it consistently. We can also assume that if you're willing to pay zillions of dollars, you could have a large display like that - produced by lucky fluke in an error-prone process - much sooner. And of course you'll get almost-as-good-but-not-retina very-high-res Android tablets sooner than that.)
Note: all of the above assumes that Apple will choose to define "retina display" as the same DPI for the iPad as for the iPhone. They probably won't. Being masters of reality distortion, I would bet on them being able to deliver their newly-redefined "retina display" iPads at least a year sooner at a lower DPI. So let's say sometime in 2012 instead. But not 2011.
By the way, if we grew the screen further, to what's now a 1600x1200 display, that's another 2.44x the area beyond the iPad, so 1.29 more doublings, which is about 2 more years. So we can have retina quality laptop or desktop monitors in, say, Christmas 2015. Since I've been wanting 200dpi monitors for most of my life, I'll be looking forward to it.
I bet "2x" mode for viewing iPhone apps on the iPad 2 will look a lot
prettier than on the iPad 1, though. After all, iPhone 4 screens already
*have* twice the pixels, so "scaling it up" by 2x isn't actually scaling it
up, it's just displaying the same pixels at a larger size. The iPad 2 will
probably pretend to be an iPhone 4 to iPhone apps.