wiki:OlderNews

Version 3 (modified by admin, 10 years ago) (diff)

--

December 9, 2011: We broke 200 GFLOPS running HPL on the a Limulus! The latest version uses the Intel i5-2400S processor. (Note: the results are double precision CPU GFLOPS.)

September 15, 2011: Once we figure out how to package and power things, we will be seeking a vendor to manufacture Limulus!. If you are interested, join the  Limulus Announce List or the Twitter feed.

March 23, 2011: We are evaluating Sandy Bridge (i5-2400S) systems now that the motherboards are fixed. There is now an announce mailing list (see above).

January 18, 2011: Software is almost done. Expect the repository to go up any day now. Check out the automatic power control test in the Limulus Software Page.

September 14, 2010: The improved case is almost done (really). Who knew sheet metal could be so much fun. Lot's of RPMS are done. I also decided on the next system to build with the new sheet metal. I'm going to use AMD quad-core 910e (2.6 GHz @ 65 Watts) for the three nodes and a six-core AMD 1055T (2.8GHz @ 125 Watts) for the head node. That will be 18 cores, one box, one power supply, cool, quiet, and fast. Should have it done in October time frame.

August 5, 2010: The improved case is almost done. We have been moving on both hardware and software fronts and expect some announcements this fall. There are also a bunch of Xeon 3400 series Micro ATX motherboards now available that support dual Intel GigE and ECC memory. (Supermicro, Asus and Gigabyte).

April 30, 2009: I have been busy with the hardware aspects of the project. The goal is to try and fit Norbert Limulus Cluster now fits in a single case measuring 23x20x8.5 inches (58x51x21 cm). It uses a single power supply, is very power efficient, quiet, and looks rather cool. There will be more real soon. The first version will use quad-core CPUs for a total of 16 cores per system. Also, there has been some concern about using non-ECC memory. Based on current memory technology, actual tests, and some research papers I do not think this is going to be a concern for most users. More information when the tests are done.

October 22, 2008: I have posted the new  results and the issue with Jumbo frames was due to flow control in the NICs. I used ethtool to turn off flow control and was able to see the expected performance boost for larger frames, the variability seemed to get worse however. I'll be writing this up and posting it on  ClusterMonkey real soon. Also,  Open MX 1.0.0 has been released.

August 13, 2008: I recently  posted some interesting benchmarks for Norbert. I have been testing  Open-MX over GigE and found some lower throughput when Jumbo frames are used. I think it could be the low cost switches. More testing later this week. I just noticed that the latest version of OpenMX (0.9.1) can run over standard frame sizes (1500). More testing indeed.

July 31, 2008: At long last an update! I have been busy with the software. Our goal is to have a Fedora 8 spin ready by the end of August 2008. We will also be adding our complete set of spec files plus our build/install scripts to the source tree. We will also host {S}RPMS at this time. Staring in September we will be working on packaging issues (i.e a case). In the mean time you can get more background from a pre-print (draft actually) of my  Limulus Article (pdf) in the November 2007 Issue of  Linux Magazine (It has not made it to the Linux Magazine website yet).

January 1, 2008: The Trac page is up and running. The next step is to get the Wiki and Milestones completed. There is now a hardware manifest and pictures for the The Norbert Limulus Cluster.