#907 1998 · Akamai Technologies (Tom Leighton, Danny Lewin) · Internet infrastructure
Akamai answered internet traffic jams with an algorithm instead of bigger servers
问题
Popular sites collapsed under 'flash crowds', sudden traffic surges, no matter how much server hardware backed them
背景
In 1995, Tim Berners-Lee challenged MIT researchers to solve a problem the web's own success was creating: whenever a site got popular fast — breaking news, a product launch, a viral link — the single server or server farm hosting it would buckle under the surge, an effect the industry called a flash crowd. The standard fix was to keep buying bigger, more central infrastructure, but that only raised the ceiling; it didn't change the fact that every request for a given page still funneled through the same physical location; and it did nothing for the network congestion on the way there.
Tom Leighton, an MIT applied-math professor working on parallel algorithms and network theory, took up the problem with graduate student Danny Lewin. Their question wasn't how to make the origin server bigger, but whether the request needed to reach the origin server at all — whether the content could instead be found near whoever was asking for it, chosen dynamically by a formula rather than by a fixed, pre-assigned location.
换别人会怎么做
The available response to a flash crowd was to over-provision the origin — buy more servers, more bandwidth, more headroom at the one location everyone's request had to reach — which raises cost roughly in proportion to peak demand and still leaves a single point of failure.
他们看到了什么
The bottleneck was not server capacity, it was routing every request to one fixed place. Once content could be replicated near the requester, congestion became a routing problem algorithms solve cheaply.
那一手
Leighton and Lewin, incorporating Akamai in August 1998 with an exclusive license to their MIT research, built a system that used consistent-hashing algorithms to replicate a site's content across thousands of servers scattered across the internet, then routed each individual user's request to whichever nearby copy was fastest and least congested at that exact moment — deciding this in real time rather than sending every request back to one origin.
为什么管用
The mechanism works because it turns a single chokepoint into thousands of interchangeable, geographically distributed options, so no individual failure or surge can take down access to the content — the consistent-hashing algorithm can reassign requests around a failed or overloaded server in real time without anyone needing to notice. Cost also scales with the network's aggregate capacity rather than with the worst-case peak at one location, so the same infrastructure serves both ordinary days and flash crowds without being sized for the rare event.
值了多少
The network absorbed billions of daily requests by rerouting around congestion; Akamai's 1999 IPO opened at $26, closing near $145 that day.
什么时候会失灵
The approach needs content that can be usefully cached or replicated near the edge; it does much less for highly personalized, constantly changing, or transaction-heavy requests that still have to reach a central system of record. It also requires a genuinely distributed physical footprint — the algorithm has nothing to route to if there aren't already servers positioned close to where demand is concentrated, which is a real infrastructure investment, not just a software trick.
后来呢
Akamai's edge-caching, distributed-routing model set the pattern every major CDN since — Cloudflare, Fastly, Amazon CloudFront — has followed, and it underlies the assumption, now taken for granted, that a website can serve a global audience without a single point of failure at its core.
资料来源
- [1]Akamai's AlgorithmsMIT Technology Review, 2000technologyreview.com