Limited availability of Dell’s PowerEdge 3250 Itanium 2-based servers meant that the company couldn’t provide a review cluster to the InfoWorld Test Center. So, InfoWorld sent me to the Dell’s new HPC lab in Austin, Texas, to examine the PowerEdge 3250 and an eight-way HPC cluster based on that server. Dell had already pre-built the servers and networks. The only tasks remaining for me were installing the software and beginning the number-crunching tests.
In effect, the node servers are like black boxes: Although Dell lab engineers installed KVM (keyboard/video/mouse) connectors on the node servers for the test, it’s unnecessary to manage the nodes directly during normal operations. Direct control of those servers would generally be required only when installing firmware upgrades or diagnosing hardware problems; the Ganglia tool provides the necessary feedback regarding each server’s performance and CPU and network loads.
The configuration I tested consisted of an eight-node cluster, with each node being one Dell PowerEdge 3250 server with two 1.3GHz Itanium 2 processors and 4GB of RAM. The cluster also contained two other PowerEdge 3250 servers, one configured as a “master node” that distributed work to the compute cluster and coordinated the efforts of all the machines, and the other as a “management node” that provided the administrative interface, as well as systems monitoring through the Ganglia application.
There were two interconnects used for the test network: Gigabit Ethernet (using the servers’ built-in NICs and a Dell switch) and Myrinet (using network cards and an eight-port switch from Myricom). Both interconnects can be specified as part of a cluster solution provided by Dell. The benefit of the Myrinet system is that its streamlined protocols and custom hardware have a latency about one tenth that of the Gigabit Ethernet, thereby eliminating any delays when one server needs to get information from another.
The testing itself consisted of deploying and running several applications on the test cluster, culminating in running an industry-standard supercomputing performance test, Linpack , which solves a complex, memory-intensive mathematical problem using sophisticated 64-bit floating-point calculations. Benchmarks such as Linpack measure the gigaflops (billion float point operations per second) of the cluster, allowing the efficiency of the cluster, or its actual versus theoretical performance, to be easily calculated.
Finally, it’s interesting to note that not all applications of a cluster would lend themselves to Itanium 2 servers such as the PowerEdge 3250; others, for example, would be served equally well by Dell’s less expensive PowerEdge 1750, a Xeon-based server the company also offers in an HPC configuration. Both servers, in theory, would perform equally. Because the Itanium 2 processor has two full-speed floating-point units within the chip, the theoretical maximum performance of a 1.5GHz processor — the fastest currently offered — would be 3.0 gigaflops.
That’s exactly the same theoretical maximum as a 3.0GHz Xeon processor, which only has one floating-point unit. However, the Itanium-based server can do 64-bit math; the Xeon, only 32-bit math. For some applications, the extra precision may be important. But in other operations, such as those that use integer math (where the Itanium 2 and Xeon are on equal footing), the fast clock speed of the Xeon chip would result in better supercomputing performance.