Credit: Google
Web
How to Make the Internet a Lot Faster
Google's promise of very-high-speed broadband can't just be about big pipes.
- Thursday, February 18, 2010
- By Erica Naone
Last week, Google announced its plans to build an experimental fiber network that would offer gigabit-per-second broadband speeds to up to 500,000 U.S. homes. Among other goals, the company said it wanted to "test new ways to build fiber networks, and to help inform and support deployments elsewhere."
Google hasn't released many details yet, but experts believe that the key to successful very-high-speed broadband doesn't lie in fiber alone. To really speed up the Internet, Google will have to operate at many levels of its infrastructure.
Gigabit-per-second speeds are much faster than, for example, the speed currently offered by high-speed services such as Verizon FiOS. However, Google's network won't be the first to reach such speeds. There are several such deployments internationally, including in Hong Kong, the Netherlands, and Australia. Internet2, a nonprofit advanced networking consortium in the United States, has been experimenting with very-high-speed Internet for more than a decade, routinely offering 10-gigabit connections to university researchers.
Existing applications for very-high-speed Internet include the transfer of very large files, streaming high-definition (and possibly 3-D) video, video conferencing, and gaming. Some experts speculate that accessing large data files and applications through the cloud may also require better broadband.
"Just big pipes alone to an end user does not necessarily guarantee that you can deliver high-end applications," says Gary Bachula, vice president of external relations for Internet2. There are many factors beyond raw bandwidth, Bachula says. For example, an improperly configured router or a university firewall can affect performance and end up acting as a network bottleneck.
"You need to have open networks, you need to publish your performance data, you need to have people troubleshoot your network remotely," says Bachula. In recent years, Internet2 has been researching tools and technologies that can help find and resolve the performance issues that occur on high-speed connections "in a systematic and seamless way." Ideally, he says, consumers as well as network managers would be able to use these tools to diagnose the network.
"If we're really going to realize the vision of some of these high-end applications, it does have to go beyond basic raw bandwidth," he adds.
It's also not enough to build a fast hardware infrastructure, says Steven Low, a professor of computer science and electrical engineering at Caltech, and cofounder of the network optimization technology company FastSoft, based in Pasadena, CA. Low believes the protocols that move traffic through the network will also need to be updated to make effective use of very-high-speed capabilities.
- 1
- 2
Related Articles
City with Superfast Internet Invites Innovators to Play
A Tennessee city with one-gigabit-per-second Internet runs a $300,000 contest to find ways of using it.
A Network for Wild Experimentation
ESNet and Internet2 have built a networking test bed for ideas that are too disruptive for the networks that people actually use.
The Slow-Motion Internet
Google's growth plans depend in part on whether it can make the entire Web faster.
Mapou
357 Comments
- 1190 Days Ago
- 02/18/2010
Nicely researched and written article. The folks at Google have got their work cut out for them. People will never be satisfied with their bandwidth because, as soon a new bandwidth capability is installed, new demanding and gargantuan services will emerge to gobble it up. Imagine a time when you can participate in full 3-D, high resolution virtual conferences that look like the real thing. Facebook would never be the same after that. 10 gigabits may sound like a lot of capacity to move bits around but the internet of the future will be so immersive and so intense (in ways that we can't even dream of) that, 10 years down the road, it will barely suffice. And consumers will clamor for more speed.
Sure, go ahead and install as much fiber as you can (is there a Moore's law of internet bandwidth growth?) but there must be a practical limit to how much fiber you can spin out. What are we going to do when that happens? Fortunately, physics may offer a solution. There are excellent reasons to suppose that space (distance) is but a perceptual illusion. Nonlocality is synonymous with nonspatiality. In the not too distant future, we'll have technology that will allow us to move matter instantly from anywhere to anywhere. This will afford us, not just desirable conveniences like teleportation and truly secure peer-to-peer networks, but as much bandwidth as we can use. It may sound like sci-fi but it's more probable than you think.
msreid
27 Comments
- 1190 Days Ago
- 02/18/2010
Yes, I believe there is an upper limit on how much we will ever need, and it has nothing to do with technology or costs or environment. It has to do with the actual capacity of the humans who use the system.
For example, we have a certain maximum resolution to our vision, and so improving a screen to give even better resolution than that would be a waste, as we cannot process more than a certain amount.
The same is true for all of our senses and capabilities. There is an upper limit to the hardware and software capabilities of our minds and our sensory organs. And when we get applications created that maximize all of that, there will be no more need to increase bandwidth on a per-user basis.
Maybe if we can find ways to upgrade our bodies' hardware and/or software, then we'd have need for a higher maximum.
ztracy
8 Comments
- 1187 Days Ago
- 02/21/2010
This is a great analysis. It seems clear that human end user requirements will lessen over time but what about the systems that our mobile phones will evolve into? What services will we demand from them?
carlhage
84 Comments
- 1190 Days Ago
- 02/18/2010
To me "The Internet" is really slow-- not because I don't have "broadband" access, because when I click on a page it takes a long time to load. When web sites are perceived to be really slow, it isn't because of bandwidth-- it's usually because of slow database access feeding horribly inefficient page composition software, and transmitting bloated html and javascript in bits and pieces.
The Technology Review web site is an example of a good site, not one of those we think of as horrible. But even TechnologyReview.com is highly inefficient. I just ran a small test on this page. The retrieval was sequential, so slower than the browser which takes about 7s to complete (thankfully the page is readable before loading everything).
Sequential retrieval for this page (before my comment bloated it) was 15.539s for 340,594 bytes in 57 files. My 1.5Mb link should take less than 2s to load if the available bandwidth were actually used. Thankfully browser tabs were invented so I can download a page on the side, then read it later when the web site finally responds. The average speed downloading this page was 175kb, or 12% utilization. I'm sure if I had one of those 50Mb services the serial access time would be faster, but the utilization would be even less.
[Note this does not include flash-- the downloader doesn't include this or run the javascript.]
The http protocol can transmit compressed data but few sites use this option. The 341K for all data on this page is 158K compressed. The main html page is 143,553B but only 26,161B compressed (82% waste). Using lynx to convert the HTML to text yields 50,107B (including all link addresses) and it compresses to 13,838B. [The 14K is the content out of 341K data exchanged. Pictures are 80544 jpg/13947 gif.] If you view the HTML you see lots of superflous space and repetitive, excessivly complex html tags-- this is quite typical. Just cleaning up the HTML and web server and transmissing compressed would reduce data transmitted by a factor of 5.
The 26K main page could appear in .14s (plus speed of light back and forth to MIT) on my 1.5Mb DSL. It appears in 4-5s on a reload and about 7s to finish all loading, so way more than 10x slower than necessary, and nowhere close to being bandwidth limited. A 1Gb link would reduce the page transmit time from .14s to 208us, but so what? It would still take 3.9s to appear!
The real way to speed up the internet is to improve the software for web servers and browsers. That is the real source of slowness. [Oh, and of course put spammers in jail.]
Of course if I were viewing a 1080p full-motion HD-video, I would need more, but I wouldn't need 1Gb.
chrisjmiller
64 Comments
- 1190 Days Ago
- 02/18/2010
While we all may covet a 1Gbps Internet link, I wonder what the real benefit would be. As the article says, existing Internet protocols (IPv4) become increasingly inefficient at these speeds, and IPv6 is little better. It's not that difficult to provide 1Gbps links within an apartment block of 100 dwellings, but aggregating these into a 100Gbps Internet connection is more challenging. And what services demand these speeds? Blu-ray movies are 54Mbps (although broadcast HDTV rates are far lower), so even without compression 3D movies can't need more than 100Mbps.
In reality, supplying a 1Gbps link to the existing Internet infrastructure is like dropping a 5-litre truck engine into a Model T and expecting it to perform like a Ferrari. The whole infrastructure would need to be upgraded in order to deliver any significant benefits to the end customer.
So what can Google offer? The current Internet is limited much more by latency - Round Trip Time, the time taken for a request to travel from your computer to the server (on the other side of the continent or even the other side of the world) and back again - than it is by data rates. I'd benefit much more from halving latency than doubling my line speed. Google can address this problem. They already have (in effect) multiple copies of the Internet within the many data centres they use to support their search facilities. Instead of my web requests crossing the globe in multiple hops, they could satisfy the majority of them from a 'local' data centre. Now that could be something worth paying for.
dtutelman
117 Comments
- 1190 Days Ago
- 02/18/2010
ChrisJMiller correctly identifies two culprits -- two challenges to making proper use of higher speeds -- as latency and protocol design. But his suggestion (duplicating the information geographically) does not make the connection between the two culprits.
Forty years ago, virtually all protocols were ACK/NAK affairs. Windowed protocols were introduced mostly because satellite communication increased the latency. Windowed protocols allow multiple unacknowledged blocks to be in transit, allowing better use of the "pipe".
Windowing addresses the ratio of block transmission rate (in "fire hose" mode) to the inverse of the latency (the rate at which acknowledgments can be round-tripped). The ways the network designer can address it, obviously, are:
* Slower transmission. No, scratch that. It increases the efficiency, true. But it reduces the overall throughput, so it's counterproductive. No progress gained with this one.
* Bigger blocks. That may help for transferring large files. But it is limited by (a) interactive traffic, which naturally has smaller blocks, and (b) error rate on the medium, which hits too-big blocks with too many retransmissions.
* Bigger windows. Yeah, that's the ticket. Allow more blocks and acknowledgements into the pipe before you have to stop and wait to transmit more data.
Windowing has been a very effective strategy in the past for dealing with both increased latency and increased bit transmission rate. The existing TCP/IP protocols have windowing; it would not take much design change to radically increase the window size. We should probably try that first. (And I know design is not the hardest part. Migration of the whole world to the new protocol is.)
Yes, more may be needed for worldwide gigabit rates. But windowing should be the first thing we try.
As for duplicating the information close to the user, the biggest problem there is not cost. Actually, the cost of storage and processing is cheaper (and getting cheap faster) than the transmission costs. The big technical challenge will be synchronizing the data, making sure everything is up-to-date. Perhaps the biggest use of gigabit speeds will be streaming media, where synchronization is not an issue. But, for many network uses, up-to-date data is important, and duplication is nontrivial -- and is itself network-intensive.
ms
190 Comments
- 1190 Days Ago
- 02/18/2010
IP has very little to do with efficiency of end-to-end connections, and it is actually other protocols that determine the route that any particular packet takes through the network. The main difference between v4 and v6 is the size of the address space, which has nothing to do with efficiency. Note that the article says nothing about IP.
Latency is already being addressed by hosting services (such as Akamai) that provide geographically dispersed servers to service requests. But for many applications, latency is not particularly important--a 50ms latency is perfectly acceptable for human interactions, and noninteractive applications (such as streaming video and audio) are even less sensitive.
chrisjmiller
64 Comments
- 1190 Days Ago
- 02/18/2010
You might want to do some more research before suggesting that the only significant difference between IPv4 and IPv6 is 32-bit vs 128-bit addressing (although this is certainly very important and is the main driver for v6 adoption). Off the top of my head: v6 supports much larger packets, there's no intermediate packet fragmentation in v6; routing is handled quite differently. There are other substantial differences to provide better support for encryption and mobile traffic, too.
All this will make for some improvement in utilisation of faster pipes, but we're still stuck with the basic inefficiency of IP. [Exercise: run a protocol analyser on your PC when accessing a web page, how many separate TCP connections does it set up? You may be surprised!] Because of this, latency is important and that's why Akamai can offer their services to large web sites - for a fee, of course. Google could do much the same, but for every web site (or, at least, most of the popular ones).
Oh, and IP is still the main determinant of Internet routing (unless you're using a private MPLS network or similar).
ms
190 Comments
- 1190 Days Ago
- 02/18/2010
Yes, v6 allows larger packet sizes, but the max v4 packet size is .5 Mb, which is .5 ms at 1 Gb/s. Packet processing is a tiny fraction of that .5 ms, so using larger packets isn't going to noticeably affect efficiency. Lack of intermediate packet fragmentation in v6 might make v6 routers slightly simpler, but again it isn't going to noticeably affect efficiency. Routing still uses BGP, whether for v4 or v6. Sure, there are other differences as well, but nothing to noticeably affect efficiency.
What does the number of TCP connections have to do with IP?
IP does not specify the routing protocol; that's up to the network designers, who have currently agreed on BGP.
chrisjmiller
64 Comments
- 1190 Days Ago
- 02/18/2010
What does TCP have to do with IP? Well, 90% of IP traffic on the internet is in the form of TCP and 90% of that amount is HTTP-related. Establishing each TCP connection requires an inital 3-way handshake, so that's already 1.5x the round trip time before any data has been exchanged. My link to Technology Review requires 15 'hops' (intermediate routers) and a RTT of 150ms, so that's quarter of a second right there.
BGP (Border Gateway Protocol) is one of several protocols that may be used to determine the routing of IP datagrams. It is used only between 'autonomous systems' and so is present only on a very small proportion of routers - probably representing one or two hops out of the 15 between my computer and the TR web server - certainly not the routing protocol on the Internet.
To avoid boring the good readers of Technology Review, I'd be happy to continue this discussion by private message.
annroberts
1 Comment
- 1190 Days Ago
- 02/18/2010
At first, you are happy about fast speed, then you wanna gat more and more. It will never be enough...
ms
190 Comments
- 1190 Days Ago
- 02/18/2010
While TCP is a 30-year-old protocol, it is continually being updated with new options. There's no reason to think it doesn't work fine at Gb/s speeds. The main problems have to do with stalling transmission while waiting for packets to be acknowledged, using bandwidth to retransmit packets that were actually received successfully, and having to wait for retransmission of lost or damaged packets. Suitable option choices can ameliorate all of these problems.
chrisjmiller
64 Comments
- 1190 Days Ago
- 02/18/2010
I would never malign TCP, and it will work fine at Gb/s speeds, but you won't get anywhere near 100x the real-life throughput compared to a 10Mb/s link. This is why folks may prefer to use other protocols, such as fiber channel, over multi-Gb links in data centres.
As I pointed out above, there are currently no end user applications that demand Gb speeds, so maybe the inefficiency is academic.
ms
190 Comments
- 1190 Days Ago
- 02/18/2010
I wasn't accusing you of maligning TCP, I was accusing Erica of doing so.
But, with proper settings, TCP can deliver throughput proportional to the raw bandwidth.
Of course, different network architectures, particularly in controlled environments, may be preferrable in the face of different (from the Internet) constraints and requirements.
GaryB
119 Comments
- 1190 Days Ago
- 02/18/2010
"there are currently no end user applications that require 1Gb"
Are you insane? It's like saying "there currently is no market for the fountain of youth". True, because it doesn't exist ... yet. I have applications that require > 1G now -- teleop assisted teleop robotics, cloud<->local host object recognition, life logging, real time language translation. I could go on and on ... think about "dumb" equipment controlled elsewhere by very smart machines, ... don't get me started on 3D printing, swappable virtual machines for games or other local dedicated tasks.
chrisjmiller
64 Comments
- 1189 Days Ago
- 02/19/2010
Which part of the word "currently" do you not understand? I stand by my view that there is no current application for ordinary home users (which is what the Google proposal is about) that requires anything more than 100Mbps. I'm sure there are large institutions that need this kind of capacity today (the LHC springs to mind) and maybe in the future there will be domestic applications as well. If we ever do need to deliver 1Gbps to every home, upgrading the central infrastructure will be more of a challenge than the 'last mile'.
joek1970
1 Comment
- 1190 Days Ago
- 02/18/2010
Streaming is still a Problem, File Transfer Problem Has Been Solved!
Latency and packet loss is a big challenge and would continue to dog streaming even if you have 1 Gbps connection.
New protocols that are independent of latency and robust to packet loss are already delivering Gbps+ file transfer speed, and have become a standard in the media and entertainment industries. Collaborative movie production teams, such as the team that developed AVATAR, depend on these new file transfer protocols for their critical transfers, often across international boundaries
see www.asperasoft.com
fiberman
186 Comments
- 1190 Days Ago
- 02/18/2010
Fiber to the home (or any user) is essential for providing greater bandwidth, and fiber has lots more tricks up its sleeve. FTTH currently uses network type multiplexing and addressing but fiber has the capability of carrying multiple signals at different wavelengths (WDM = wavelength division multiplexing) as is done on most long distance networks today. FTTH passive optical network (PON) systems are already being tested where each user has their own wavelength (WDM PON) upping bandwidth by 32 to 128 times on the same fiber.
The primary requirement to creating future capability in the terabit range is to get the fiber connection direct to the end user, not kludge a system that still uses copper as the final connection to the user!
Have you seen this: http://www.ntia.doc.gov/reports/2010/NTIA_internet_use_report_Feb2010.pdf
Lots of people still don't want or want to pay for broadband!
fiberman
186 Comments
- 1190 Days Ago
- 02/18/2010
I've been around in this biz for over 30 years and participated in the transition to fiber and digital. One thing I never understood - why everybody wants to force everything into one box - or in this case protocol.
Voice, data and video have very different transmission needs. POTS worked great for voice, but once digitized had high overhead for data - remember when everybody was hyping ATM for data in the 1990 time frame until they realized it had over 30% overhead due to its optimization for voice. Voice requires low bandwidth so VoIP got crammed into data packets but had latency problems until bandwidth grew. Now we're trying to put video onto data and it's just as big a kludge. Video needs to stream, does not need the kinds of overhead and protection that data requires, but that hasn't stopped the stampede to overload data networks with video just like we overloaded digital phones with data a decade ago.
When will they ever learn, when will they ever learn.....
Packman17
5 Comments
- 1186 Days Ago
- 02/22/2010
If You Build It They Will Come
As we get caught up in the technology of delivering more "highway" in the form of greater throughput and speed, I think it is useful to look at analogies from other areas of society. Forgive me for not having the name of the study at hand, but a study sponsored by the UK to determine the effects on motor vehicle congestion of building a new highway, and later confirmed by reality, showed that "if you build it, they will come". If I remember correctly, less than a month after building the new highway loop, which provided a very significant amount of new road space, the new highway was jammed as bad with traffic as the old one. Apologies for the cliche, but the message appears to be a universal truth. If you provide a new and/or better highway, the volume of "traffic" will expand to utilize all of the improvement in throughput and more. Humans have an innate ability to think of new uses for available capacity of any kind. Known as "the law of unintended consequences" it has proven itself out in all areas of our society, including the current internet, where even "Al Gore" could not have imagined the uses to which it has been put once public acess became freely available.
roygbiv4242
1 Comment
- 1186 Days Ago
- 02/22/2010
what about feeble home networks?
I am amazed that everyone has overlooked the home use/network environment. How many homes now or in the next ten years would even be able to use a 1 Gb/s connection? I'll bet a year's salary that less than 0.1% of households have an internal fiber network or even Gb copper LAN (via Cat5e or Cat6) running throughout their home. And even if they did, people love their in-home wireless networks which are lucky to achieve TCP/IP throughput rates of 10 Mb/s in anything close to a suburban environment. People also love those ubiquitous cheap wireless routers with their processors which are incapable of processing 100 Mb data rates much less Gb data rates. Then you have the whole problem of the end-use appliances themselves. How many appliances will actually be able to inhale data at a Gb/s? Until the average customer is willing to spend several hundred or even thousands of dollars to build AND MAINTAIN a high-performance network environment in their home, what good are Gb/s connections to the home?
aymandaoud
1 Comment
- 1185 Days Ago
- 02/23/2010
I think google is working right now on a new set of services to become "No.One" in cloud computing, to reach this point they have to develop well infrastructure around them starts with 1G, specially when they start offering education suite for public,we can imagine how many schools and universities can handle this kind of net connection taking in account that google will advise them "their users or potential users" to stop spending their money on developing their own infrastructure"paying for new servers, blades and so on", 1G will enhance their opportunities to be Number One in the new Cloud computing world
On Twitter
Become a Fan on Facebook
Subscribe to the Feed
jmaximus9
86 Comments
Thumbs up for G
I really hope they can pull this off, I am tired of being overcharged by Comcast for last decades speed. While Japan and Korea have 40 and 60 megabit, I struggle to get 6 and have monthly bandwidth caps to boot. My speed hasn't increased in the last 10 years, but my bill sure has.
Reply
Viv
62 Comments
Re: Thumbs up for G
Yes your right, whats offered stays the same and the bills go up, a good business model for the incumbent providers with no real competition to worry about
Reply