Hello Argo, Goodbye Borg

Since 1998, the mainstay of high-performance computing (HPC) at the ACCC has been the borg server, an Hewlett Packard V-class computer. Borg was home to a diverse group of computationally intensive software packages and libraries, including Gaussian 98, ANSYS, IMSL, SAS, and SPSS. After a series of upgrades and one replacement, the ACCC began to think about replacing borg again, this time perhaps with a different type of machine.

Two factors dictated the direction the ACCC would take. One, the high cost to upgrade or replace borg with another shared-memory, multiprocessor machine as well as the fixed hardware and software charges for that type of machine. Two, the inverse relationship between the increasing performance of PC-based machines and their declining cost. The ACCC elected to replace borg with a PC-based multicomputer system known as a "Beowulf cluster."

The new system is named argo after the ship used by Jason and the Argonauts who, according to Greek mythology, embarked on a perilous journey to find the Golden Fleece.

Unlike borg, which is a single machine with multiple CPUs all sharing the same memory, argo is a multicomputer system with distributed memory; it is currently is composed of nineteen PCs. Each computer has a Athlon 1.4 GHz processor running Linux OS with 768 MB of DDR/EEC memory, a 40 GB hard disk, and an Fast Ethernet NIC. The machines are linked by both hardware and software connections that permit the separate machines, also called nodes, to function as a whole.

Users log into a single entry point, the PC known as the master node. All jobs submissions -- programs written by users or jobs using software packages -- are initiated for batch processing from the master node. The batch system sends those jobs to one or more worker nodes for execution. Worker nodes have been configured to run multiple jobs simultaneously (multitasking).

The architecture of the cluster allows for obsolete nodes to be removed and replaced with more powerful machines, or new nodes may be added without decommissioning the current ones. A node that is not functioning can be removed without bringing the entire cluster down.

In the coming months, 16 additional worker nodes, machines running 2.8 GHz Xeon processors, will be phased in. Another feature of the ACCC cluster is the internal high-speed Dolphin/Scali interconnect, which is used to satisfy the heavy network demands involved in parallel programming. Parallel programming permits a large program to be divided into smaller units with each of those smaller sections executing at the same time on different processors (nodes).

All of the packages on borg are being evaluated for transfer. Heavily used applications such as MPI, Gaussian 98, and IMSL are already installed on argo; other applications, particularly those available on other ACCC platforms as well as those not suitable for a distributed memory system, will be left behind. The plan is to complete the transfer of software and clients by the start of the fall semester.

For more information, go to the argo home page:
http: //www.accc.uic.edu/hardware/argo/

To request an account on argo, send email to systems@uic.edu. Include your name, netid, and a brief description of how you’ll be using argo.

Comments are welcome; send them to
Michael Homa, systems@uic.edu