Cluster Computing

Researchers: Ari Lukkarinen and Arto Selonen

Recent years have brought a name "Beowulf" into existence from ancient English mythology. In a poem Beowulf was a heroic prince of the Geats of southern Sweden. There is no evidence of historical person Beowulf, but Beowulfs have found their place in modern scientific community, see www.beowulf.org. Instead of poetry the name Beowulf is nowadays more often connected to terms like cluster and parallel computing, COTS (commodity-off-the-shelf) products, and Linux operating system.

The motivation for Beowulfs, or the Beowulf class of computers, was the lack of money and the everlasting need for computational power. The idea behind Beowulfs was to create an effective "virtual computer" consisting of many personal computers. These personal computers are build from commodity-off-the-shelf parts and, because of large and fiercely competed consumer product markets, these computers are cheap in a sense of computational power per money.

We entered the world of cluster computing about two years ago. Instead of relying on widely used Intel/AMD based PCs we decided to build our cluster on the basis of Alpha processors. The first cluster consisted initially of 8 computers (nodes), but it was later enlarged to 16 nodes. The next cluster was a small cluster including only four nodes. This machine was used to test the possibility of speeding up data throughput between the nodes by using multiple Ethernet connections for each node (channel bonding). The next cluster differed from previous ones in a sense that it consisted of four two processor nodes. The experience gained from all these clusters was used when we acquired together with CSC and Laboratory of Physics at HUT a large Beowulf cluster Clux consisting of 96 nodes. Usually nodes use inexpensive Ethernet network and message passing library (MPI/PVM) to communicate between each other. However, some applications benefit a lot from faster networks made specifically for clusters. This is why our cluster has 64 nodes interconnected with a very fast and a low latency Myrinet network. The rest of the nodes use a Fast Ethernet connection. Cluxi was put into public use in November 2000. Since then we acquired a 16 node version of Clux and also an eight node Pentium 4 (2 GHz) cluster.

Figure 42
Figure 42:1/8 of the "Clux" cluster.

www@lce.hut.fi