내 삶의 Variation

EVER noticed how long it takes for web pages to load these days? You click on a link and wait and wait, and then wait some more, for the content to trickle in. If nothing has happened after ten seconds or so, your impatient correspondent hits the browser’s stop button followed by the reload key. In desperation, he sometimes loads the link into a second or even a third browser tab as well, and bombards the website’s server with multiple requests for the page. If that fails, he gives up in disgust and reads a newspaper instead.

Back in the early days of the internet, when most web users relied on dial-up connections, browsers were crude and web graphics were clumsy GIF files, eight seconds was considered the maximum people would stick around for a page to load. To increase “stickiness”, web designers pared their HTML code to the bone, collated their style-sheet data and JavaScripts into single files for more efficient caching elsewhere on the web, used fewer graphics and embraced the PNG and JPEG picture formats, with their smaller file sizes, as soon as they became available. Compared with text, pictures really were the equivalent of 1,000 words, at least when it came to the time taken to transmit them.

When your correspondent hand-coded The Economist’s first website back in 1994, a typical web page was about 50 kilobytes in size and dial-up modems could transfer no more than three kilobytes a second. To stay under the “eight-second rule”, pictures were kept to a minimum, so no page took more than three or four seconds to begin loading and never longer than 20 seconds to complete. The irony is that, with broadband nowadays more or less everywhere, overall connection speeds have gone up by leaps and bounds, yet the time taken to load web pages seems only to have got longer.

Your correspondent is admittedly near the end of the road for a digital subscriber line (DSL) connection. But even at three miles (5km) from the local telephone exchange, the speed of his broadband connection has inched up over the past few years from 65 kilobytes a second to more than 90 kilobytes a second—as the local line has been tweaked and legacy equipment like echo-cancelling coils removed from its junction boxes.

Sure, he could get 650 kilobytes a second or more from a cable connection. But that would mean ditching his otherwise excellent satellite-TV service. Besides, optical fibre is slowly working its way up his hillside. He could soon have access to the internet at more than six megabytes a second—providing he is prepared to pay $140 a month instead of $21 for his existing DSL connection.

A 70-fold increase in speed for a sevenfold increase in price would seem a bargain. But your correspondent is not sure that more raw speed will solve the glacial loading problem. Even with his wimpy DSL connection, pages are rendered quickly enough once the website’s servers (and all the other computers along the route, plus those used to host adverts, graphics and miscellaneous layout bits) start giving his browser’s request some attention. The trouble is getting their attention in the first place.

Before two computers can exchange information, they have to agree to talk to one another. Under normal conditions, this requires the user’s computer to send a request to the host computer, which then sends a response back to the user. Only after this “handshaking” is complete can the exchange of data commence. The time taken for this round-trip of request and acknowledgment determines the network’s latency.

The latency cannot be less than the distance the electromagnetic signal has to travel divided by the speed of light. For instance, your correspondent’s home in Los Angeles is 400 miles from a colleague’s in San Francisco. In theory, then, the shortest round-trip between the two locations is 4.3 milliseconds. But if you “ping” the other computer, you’ll get a round-trip time of typically 700 milliseconds. That is still pretty quick, but it shows just how much time is spent waiting around for the various servers involved to handle the request.

There are many places along the way where the message can get bogged down. Queues can build up at routing servers that switch data packages along different routes to their destinations depending on the traffic. Worst of all, the DNS (Domain Name Server) computers used by your ISP can be overwhelmed as they try to translate the names of all the websites subscribers want to visit (say, www.economist.com) into their actual internet addresses (216.35.68.215). If you know it, try using the website’s numerical address rather than its verbose URL (Universal Resource Locator) name. That can sometimes halve the response time.

The bottlenecks—whether at the DNS translators, the routing computers or the host’s own servers—stem largely from the way the mix of internet traffic has changed faster than the infrastructure used to carry it. Websites that were once just 50 kilobytes of text and tiny pictures now come with music, video and animated graphics. YouTube, Hulu, iTunes and BitTorrent have much to answer for.

It is even worse on the mobile phone companies’ proprietary networks. Carriers are struggling to keep up with demand as subscribers use their smart-phones to check Facebook, stream videos from YouTube and play interactive games. Where a mid-range smart-phone would consume about 100 megabytes of data a month, more advanced models like the Apple iPhone or Motorola Droid, with fully fledged browsers and access to thousands of downloadable applications, tend to consume over 500 megabytes a month. With the imminent arrival of tablet computers like the iPad, which come with wireless modems, the appetite for downloadable data could hit a gigabyte a month (see the lead story in this week’s Business section).

And this is just the beginning. On the internet, the average latency for corporate websites in America is currently around 350ms, according to the Network Weather Report operated by the University of California, Los Angeles. Google’s latency is 150ms, Facebook’s 285ms and YouTube’s 515ms. Such latencies will have to come down considerably if the next generation of internet applications, such as telepresence, high-definition video streaming and remote surgery, are to fulfil their promise.

The future is beckoning. Netflix has just announced an on-demand video-streaming service offering full high-definition picture quality (so-called 1080p, which has 1,080 lines in its picture) with 5.1-channel surround sound. Each stream being watched will require a megabyte a second of bandwidth and a latency of less than 60ms if it is to deliver crisp, pin-sharp video and pristine sound.

For the internet service providers, that means stepping up investment substantially. But adding a lot more routers to the internet would complicate matters hugely and do little to solve the latency problem. If anything, it would actually increase the number of potential bottlenecks.

A better solution might be to light up more of the “dark fibre” installed during the heady days of the dotcom boom, but left lying unused beneath the streets since the bubble burst nearly a decade ago. That is what a number of securities firms have been quietly doing. When shaving a millisecond off the time needed to execute automated trades can increase revenue by $100m, there is plenty of incentive to build private optical networks with latencies approaching zero.

Indeed, Google said this week that it was not going to hang around waiting for the telecoms industry to build the new optical web. The company is planning a low-latency fibre network that will be capable of delivering speeds of over 100 megabytes a second for communities of 50,000-500,000 people. With luck, other internet service providers everywhere will get the message.