Xserve G5 Compute Cluster

Built in the spring/summer of 2004, our compute cluster contains 8 Apple Xserve G5 cluster nodes with dual 2.0 GHz Power PC processors and 1.0 GB of memory per node. The total available compute power is either 16 processors with 500 MB of memory each or 8 processors with the full 1.0 GB of memory allocated to each processor. The nodes are presently connected using the built-in Gigabit Ethernet and an Asante Friendlynet gigabit workgroup switch, while the possibility of using Myricom 16-port Myrinet switch looms on the horizon. Power is conditioned and supplied to the system via an APC Smart UPS 3000 power supply and battery backup. The cluser nodes, power supply and switch are housed in an XtremeMac Xrack Pro rackmount enclosure.

Functioning as the head node, running jobs and managing the cluser is a dual 1.8 GHz PowerMacintosh G5. This system is running Mac OS X Server with 2.5 GB of memory. A scratch space of 234 GB can be utilized by researchers to accomodate large data sets and output files. Currently, parallel computing is managed via the implementation of LAM/MPI 7.0.6, a high-quality open-source implementation of the Message Passing Interface specification (MPI). C/C++ and Fortran compilation is possible here using IBM's XL compilers, with a number of optimization options specifically tuned for the G5's PowerPC microprocessor. A variety of server tools are used to monitor the system's performance and usage, including Apple's Remote Desktop, Server Monitor, Workgroup Manger and Server Admin.

Benchmarks have been carried out comparing the G5 to other commercial platforms (see below), and the results show incredible speed for the relatively small cost. Similar implementation of the G5 for High Performance Computing (HPC) has been most recently made famous by Virginia Tech's Project X, a supercomputing cluster originally comprised of 1,100 PowerMac G5's, which now has been upgraded to Xserve G5's similar to the ones used in our cluster. This supercomputer is capable of world-class processing power, at a fraction of the cost of commercial supercomputing sytems, a scaled-down version of which we have implemented here in our lab.

 

Benchmark Data Slideshow

Cluster Pictures