MIT researchers have developed a new system for efficiently allocating bandwidth in data center networks. The system, which would make all web pages components load faster, was presented at the Usenix Symposium on NSDI ’17.
Yet still, when we enter an URL, for example, the browser doesn’t know in advance the nature of the page that it will load. It will retrieve the components of the page one by one, depending on the bandwidth allocated for each.
Equal Bandwidth Among Webpage Components
Each component of a web page is dependent on a different program running on a different server in the datacenter. In addition to the processing time, the bandwidth of the network will not be allocated in an equitable way among the different programs. A page cannot be sent to the user without all of its components being ready, and if a single program doesn’t get adequate bandwidth, the entire page could be blocked.
You may have subconsciously recognized this phenomenon and simply refresh the page if you see it isn’t loading quickly.
At MIT’s CSAIL, researchers have developed a new system for bandwidth allocation in data centers.
The system, called Flowtune, is a centralized bandwidth allocator that would optimize web page load and processing time through a fair bandwidth allocation. In tests, the system showed promising results, maintaining throughput (same overall data transmission rate), dividing bandwidth more equitably among all components of a page, loading pages up to four times as quickly.
Flowtune was presented at the Usenix Security Symposium NSDI ’17 (Networked Systems Design and Implementation), held from March 27th-29th in Boston, MA.
Need for Web Speed
Browsers try to speed up the loading of pages, notably by downloading compressed resources and thus reducing the time needed to load them. But this approach is basic, and not always efficient because the browser still wastes valuable seconds deciding in what order it should load which parts.
Last year, CSAIL also developed another system, called Polaris, which reduces the load time of websites by up to 34%. To fetch resources more efficiently, Polaris analyzes each web page and creates a map-like profile, listing not only all necessary components, but also their interdependencies.
Latency is measured in milliseconds, seconds at worst, but the economic consequences of lag are serious. There is a direct link between web page load time and the turnover of e-commerce sites. Amazon reported that a latency of 100 milliseconds causes it to lose 1% of its sales, and a slowdown of one second could mean the loss of $1.6 billion in sales each year.