Technology Review - Published By MIT
Advertisement

A Faster Way to the Cloud

Amazon's new protocol should make accessing the cloud faster and more reliable.

By Duncan Graham-Rowe

Friday, September 11, 2009

smaller text tool iconmedium text tool iconlarger text tool icon

Cloud computing offers a cheap way to carry out data-intensive computing, allowing companies to effectively lease processing power from an online supplier. But uploading large amounts of data to cloud computing systems has remained costly and time-consuming.

Credit: Technology Review

Today, Amazon announced a new, ultrafast file transfer protocol designed to make uploading to its cloud service easier. The move could broaden the appeal of cloud computing by allowing smaller organizations and even individuals to upload data without expensive infrastructure.

"The biggest bottleneck in cloud computing is without a doubt the data transmission--uploading and downloading data to and from the cloud," says Ian Sommerville, of the Co-laboratory for Cloud Computing at the University of St. Andrews, Scotland. Small businesses often must choose between enduring slow data transfer rates or investing in extra infrastructure, Sommerville says.

The heart of the problem is the way one of the Internet's core features--the transmission control protocol (TCP)--works. TCP regulates the flow of data by breaking it up into small packets of information, sending each packet, and then waiting for an acknowledgment that the packet has been received before sending the next one. If a packet does not arrive, TCP either resends it or assumes that the network is being overloaded and initiates an aggressive congestion-control strategy, slowing the data rate down to avoid triggering a network collapse.

While TCP works fine for sending relatively small amounts of data over small distances, it can cause major headaches for cloud-computing customers. The distance that data has to travel--measured geographically as well as by the number of network nodes it has to pass through--affects the number of errors that creep into the signal. For example, transferring data across the United States on a 100-megabits-per-second (Mbps) Internet link can result in a latency of 100 milliseconds and a loss of about 1 percent of packets, which translates to real transfer rates of just 10 Mbps or less.

Nick Trigg of Constellation Technologies, a cloud-computing company spun out of the Rutherford Appleton Laboratory, in Oxfordshire, U.K., and CERN, in Geneva, Switzerland, says that TCP can be a dramatic bottleneck for large amounts of data. This means that sometimes it is faster to physically deliver data on a disk than to upload it, he says.

Story continues below


To solve this problem, Amazon Web Services will use technology developed by Aspera, based in Emeryville, CA, called the Fast And Secure Protocol (FASP).

"Our core technology is an alternate bulk data moving protocol," says Michelle Munson, Aspera's CEO and cofounder. "The inefficiency [with TCP] is really very noticeable when transferring large amounts of data," she says.

Comments

  • cloud computing
    With the ever increasing data uploading & transmission "cloud computing " will play an decesive role in the Internet  world.
    Rate this comment: 12345

    andy_mydur
    09/11/2009
    Posts:1
    Avg Rating:
    2/5
  • Image
    It's unfortunate that the image of a tower of binary numbers terminating in a cloud of smoke was chosen for page 1 of this story on 9.11.2009.  On any other day I wouldn't think twice about it, but today it turns my stomach.
    Rate this comment: 12345

    tvdoherty@gm...
    09/11/2009
    Posts:1
    Avg Rating:
    3/5
  • [no subject]
    Would this be applicable to torrenting? If everyone used this protocol could it increase upload and download speeds?
    Rate this comment: 12345

    mtarabul
    09/11/2009
    Posts:2
    Avg Rating:
    5/5
  • TCP misinformation
    The description of TCP in this article is misleading. While it is true that you can make TCP behave as described, that is by no means the only option. TCP provides for negotiable packet size, a negotiable size for a sliding window of outstanding (unacknowledged) packets, and also provides for selective acknowledgment of packets. This means that TCP can efficiently utilize a large bandwidth with significant latency and packet loss. It pipelines packet transmissions and only retransmits lost packets. There are issues with TCP that might require a different protocol (usually based on UDP), but they don't appear to be described in the article.
    Rate this comment: 12345

    ms
    09/12/2009
    Posts:141
    Avg Rating:
    4/5
    • Re: TCP misinformation
      While it is true that TCP can be implemented to achieve very high transfer rates, such implementations are not conformant to the standard and are usually deemed bad network citizens.  The reason has to do with congestion control.

      TCP has a behavior built into it called "slow start," which causes TCP senders to stop and wait for each one or two packets to be acknowledged when it starts up or, after stabilizing, it detects a missing acknowledgment.  After dropping into slow start, the sender will rapidly increase the amount of data sent before requiring an acknowledgment, then attempt to stabilize the amount of data sent at a size appropriate for avoiding congestion.  (I'm oversimplifying this a bit...)  The net result is that TCP is intentionally very sensitive to packet loss, which is usually a very good indication of network congestion.

      Slow start was added to TCP to prevent it from congesting intervening network routers, and then continuing to overload them until they underwent congestive collapse.  In essence, slow start causes any individual connection to behave suboptimally, but it tends to make transfer over the internet as a whole, with tens of thousands of TCP connections, as efficient as possible. 

      Congestion control is very, very tricky.  I wonder if FASP has been engineered adequately to avoid congestion problems.  If not, you'll see a lot of pushback from tier 1 and 2 ISPs and FASP will go the way of the other twenty or thirty attempts to engineer something better than TCP.
      Rate this comment: 12345

      theradicalmo...
      09/16/2009
      Posts:13
      Avg Rating:
      4/5

Log In

Forgot your password?     Register »
Advertisement
Technology Review January/February 2010

Current Issue

Security in the Ether
Information technology's next grand challenge will be to secure the cloud--and prove we can trust it.
Advertisement
Advertisement
Subscribe to Technology Review's daily e-mail update. Enter your e-mail address

TECHNOLOGY RESOURCES
Advertisement
MIT Massachusetts Institute of Technology © 2010 Technology Review. All Rights Reserved.