Recently, Los Alamos National Laboratory in the US unveiled a compact, "Bladed Beowulf" supercomputer it dubbed Green Destiny.
Green Destiny is made up of 240 RLX Technologies ServerBlade blade servers configured in a Beowulf cluster, in what may be the first use of the ultradense bladed architecture for a supercomputing cluster.
Los Alamos says it learned many lessons from implementing bladed servers - primarily, that supercomputers such as Bladed Beowulf shouldn't be looked at simply on a performance basis anymore, but rather from a price/performance point of view. I'd have to agree.
Wu-Chun Feng, originator of the project, said that - even though tradeoffs were made for overall CPU performance and internal bandwidth between components - because the servers consume less power, they will be easier to maintain.
Los Alamos built the cluster by ganging up RLX servers, which use a low-power Transmeta chipset operating at 633 MHz.
There are two possible deal-breakers to consider in clustering server blades in the fashion Los Alamos did: the provisioning software and the server interconnect. In the Los Alamos case, Feng used the 100M bit/sec Ethernet adapters supplied with the RLX blades connected to a 10/100M bit/sec Ethernet switch. For provisioning software, he used RLX's Control Tower 2.
Yeng says that Los Alamos' Little Blue Penguin, a VALinux cluster of 128 CPUs took "days (and arguably, weeks) to install, integrate, and configure properly and initially required daily intervention and maintenance by the technical staff."
Feng factored the system administration cost - the installation, configuration, upgrading, maintenance and support of the servers - into his return-on-investment calculations. At $100 per hour, previous configurations would cost as much as $15,000 per year to maintain. With the RLX software, the software for each server was installed and configured in just over two hours, costing a mere $200 per year.
As for the server interconnect, some observers would say that Ethernet is not sufficient. HP has proposed using InfiniBand in its second-generation blade server architecture. Dell will implement the much faster 2.5G bit/sec InfiniBand in its blade servers.
InfiniBand has not arrived yet, though. Vernon Turner, an analyst for IDC, says that by 2005, 50% of servers will be InfiniBand-enabled. Until then, if we want a faster interconnect than Ethernet, we have a lot of proprietary cluster interconnects such as Myricom's Myrinet to chose from.