Home
Learning About Clusters
Programming Clusters
Administering Clusters
Benchmarking Clusters
File Systems for Clusters
Cluster Applications/Grid
Cluster News
Site Map
 
    Home
Search
Monkey Support
Main Menu
Home
News
Features
Columns
Reviews
Links
FAQ's
Contact
Site Information
Cluster Classifieds
Projects
Conference Reports
Cluster Agenda
Site Map
ClusterRant
Add This Article
Login Form





Lost Password?
No account yet? Register
Syndicate

Visit Basement Supercomputing

Cluster Builder

Cluster Courses


Ethernet Breakthrough Demonstrated Print E-mail
Written by Dan Stanzione   
Sunday, 01 April 2007

From the Ethernet rules the day, department

We are pleased to announce the new software package developed by a team of Cluster Monkey researchers including Robert Brown, Jeff Layton, and Douglas Eadine. The project was sponsored by the Beowulf Legacy Enhanced Ethernet Project (BLEEP). The results of the these efforts is a new software package called Geyser™ that will be available for Linux later this month. By substituting the Geyser™ drivers in the Linux Kernel, a standard Ethernet card can transmit at more than double the standard Gigabit Ethernet rate. Moreover the single bit latencies are less then one microsecond. Please read the theoretical background to fully understand this report.

The driver is open source and will be available later this month after the patent is issued. Using a combination of mathematical filtering and artificial intelligence the Geyser™ drivers allow a standard Ethernet card to achieve rates that now are making many question the need for Infiniband or 10-GigE. The following results indicate the level of performance:

single bit latency .5 microseconds
maximum throughput 2143 Mbits/second

According to lead developer Douglas Eadline, the drivers work as follows. "As we all know computers transmit in binary. So when a packet is sent from one computer to another there is a series of ones and zeros. Our driver applies a Riemann filter to the stream and then uses Artificial Intelligence (AI) to predict the next bits in the sequence. As the interconnect learns about your communication pattern it becomes faster each time you send data. The maximum we have achieved is 2143 Mbits/second on standard Ethernet cards, but with proper training, I believe it could go as high as 5000 Mbits/second."

Co-inventor Jeff Layton commented as well, "Doug and I have worked on this for three years. There were many people who told us we were crazy, but we just decided to try it and see what happened. It is a good thing we did not listen to the detractors, what the BLEEP did they know".

Finally, co-inventor Robert Brown has this to say "First, let me say, Doug and Jeff really delivered. When they first described it too me I thought they were crazy. When I finally understood the technology, I jumped in to write the documentation. Here is the interesting part, once I finished the 4 GByte doc file, I sent it to both Doug and Jeff using the new Geyser™ drivers. At that stage, the Geyser™ drivers were still learning, but the second time I sent the exact same file, it was delivered in half the time. Then on April first of this year, I sent it for a third time, it took less than 30 seconds for the exact same file! Finally, I changed a single typo and resent the file. As expected it showed up at normal Gigabit speeds as the drivers began to learn about the file, but the next time I sent the exact same file, it arrived in half the time. Amazing!"

Update: Jeff Squyres also wanted everyone to know there will be an Open MPI - Geyser™ port finished in May.

Check back later this month for more news on our breakthrough!

Comment on this article
You must login to leave comments...


Other Visitors Comments
There are no comments currently....
Last Updated ( Wednesday, 14 November 2007 )
 
< Prev Article   Next Article >
Mellanox
HPC Community: Open Software, Help, and Tips
Poll
How Long Do Your Applications Run ? (approximately)
 
Latest Stories/News
Popular
Cluster Ranting By Eadline
InsideHPC
  • The Week in Vis

    Randall Hand from VizWorld.com, the web's best site dedicated to computer graphics and scientific visualization, recap's the week's best stories related to supercomputing in the visualization and graphics industries. This week he talks about the use of simulations in Formula 1, shattering objects, and bringing down buildings.

  • Sun Video Presentation: Performance Tuning

    The Sun HPC Watercooler posted yet another helpful video presentation focused on the trials and tribulations of HPC.  This video is actually the first ‘module’ in a series entitled “An Introduction to Parallel Programming.’  The series will focus on the basics of parallel programming, debugging and general application development tips. In order to help developers and [...]

  • Green HPC podcast series, the transcript

    Just a quick note to update you on the Green HPC podcast series, which has gotten a tremendous response (thanks!). I’ve added a transcript of the first episode, in case you’re more of a reading person than a listening person. If you don’t know about the series yet, take a listen to the first episode. Get [...]

Who's Online
We have 6 guests online
Worldwide Front Page Visits

Locations of visitors to this page

Monkey Stats
Google PageRank modul - Camelpark SEO centrum

 

Creative Commons License
  ©2005-2008 Seagrove LLC, Some rights reserved. Except where otherwise noted, this site is licensed under a
Creative Commons Attribution-NonCommercial-ShareAlike 2.5 License.
Cluster Monkey Logo and Monkey Character are Trademarks of Seagrove LLC.