The TranSend Service

What is TranSend?

• TranSend is transformational because it transforms web data into a representation that is better suited to the needs and constraints of the end user. If the end user is surfing the web over a slow modem or wireless link, then the large images that are often found on the web dominate the download time of web content. A recent study performed at Berkeley showed that although most items fetched from the web are small (around 1KB), most bytes transferred to users come from large images (10KB and up). This is exactly the worst situation - the user ends up wasting most of her time waiting for a small number of large images.

  Alternatively, if the user is surfing the web from a constrained device such as a Personal Digital Assistant (PDA), most pages and images are too large to fit on a screen, and may even contain content that the device cannot render (such as progressive JPEG data).

  TranSend eliminates these difficulties by transforming data to a more reasonable form - by throwing away image resolution or color information and by transcoding data (images or text) into formats the end device can understand.

  For example, the following 39.3 KB GIF image of Vancouver, Canada can be found on the web:

  

  Over a 28.8 modem, this image would take at least 11 seconds to download; over a 14.4 modem, it would take 22 seconds or more. The TranSend proxy takes approximately 30 milliseconds to transform this image; the default transformation (which converts the GIF into a JPEG, throws away half the resolution, and turns the "quality" knob of the JPEG down to 75%) results in the following image:

  

  The transformed image is of slightly poorer quality, but is only 3.9 KB in size. This translates into a 1 second download time over a 28.8 modem (2 seconds over 14.4), or a factor of 10 improvement in speed!

  Users can explicitly control the aggressiveness of this transformation (which we call "distillation"), and can also retrieve untransformed data if they want. In this way, TranSend gives users the opportunity to spend a small amount of time getting reasonable representations of data, and then make an educated decision about whether or not to spend additional time retrieving the high quality data.

  Examples of different transformation aggressiveness for the following image of an eagle:

  	10.3 KB original		5.6 KB version
  	3.5 KB version		2.5 KB version
  	2.0 KB version		1.6 KB version
  	1.1 KB version		0.8 KB (unrecognizable) version

• TranSend is a proxy because it sits in between users' web browsers and the web servers on the Internet. When a user tells her web browser to fetch a page, the browser contacts TranSend and acts it for the page (instead of going directly to the server). TranSend then fetches the data (either out of a local cache or from the server itself), transforms the data based on a user preference profile that the user has created of herself, and returns the transformed data to the user in lieu of the original data. It is this "middleman" nature that makes TranSend a web proxy.

• TranSend is a service because this on-demand transformation is available to a large number of Berkeley students, staff, and faculty 24 hours a day, 7 days a week. Being up and running 24x7 is called being highly available; this availability is achieved through fault tolerance methods - if any component of the TranSend service fails (such as would happen if one of the machines on which it is running crashes), TranSend adapts to this failure either by shunting work away from the crashed components by attempting to automatically restart the component. TranSend is also scalable, which means if the demand for the service increases, we can add more resources (CPUs, disks, memory, or network) to TranSend, growing the service to meet the demands.

How much benefit do I get from TranSend?

Who created TranSend?

• Armando Fox / fox@cs.berkeley.edu
• Steve Gribble / gribble@cs.berkeley.edu
• Yatin Chawathe / yatin@cs.berkeley.edu
• Research advisor:
  Professor Eric Brewer / brewer@cs.berkeley.edu

• Anthony Polito / aspolito@csua.berkeley.edu
• Ben Ling / bling@cs.berkeley.edu
• Andrew Huang / achuang@cs.berkeley.edu

Daedalus and GloMop are funded by DARPA (the Defense Advanced Research Projects Agency), Hughes Research Labs, Daimler-Benz Research,

For Further Information