Linux NetworX Utilizes API Technology to Provide High Performance Computer Clusters
Linux NetworX, a Utah-based builder of Linux cluster systems, is using API Alpha motherboards and processors as the compute nodes in high performance clusters, delivering supercomputer-class floating-point throughput at a fraction of the price of proprietary machines.
CONCORD, Mass. — August 7, 2000 — Alpha Processor, Inc., through its API NetWorks business unit, today announced that Linux NetworX has selected API Alpha technology as a foundation for its high performance computer clusters. Linux NetworX designs, integrates and supports turnkey Linux clusters for universities, government laboratories and commercial engineering groups. By combining API dual-processor Alpha boards with commodity networking and its own cluster management tools, the company can offer customers a system that behaves like a single large computer for parallel workloads, yet is assembled from standard, replaceable parts.
The announcement reflects a shift that was well under way in 2000: organizations that once had no choice but to buy a monolithic vector or shared-memory supercomputer were discovering that racks of inexpensive nodes, tied together by fast networks and running Linux, could solve many of the same problems. What those customers needed was a node that delivered serious 64-bit floating-point performance and memory bandwidth, and that is precisely where the Alpha 21264 was strongest.
Why Alpha for clusters
Cluster performance is usually limited by the slowest part of the chain, and for scientific codes that is often the processor's ability to keep its floating-point units fed. The Alpha 21264 was designed around that problem. Its out-of-order core can issue up to four integer and two floating-point instructions per cycle, it has large 64 KB first-level instruction and data caches, and it connects to a dedicated off-chip backside cache that is several megabytes in size. For dense linear algebra, computational fluid dynamics, seismic processing and molecular modeling, that combination translates directly into shorter run times per node.
API's role is to make that silicon usable in volume. Instead of buying complete workstations, a cluster integrator such as Linux NetworX can buy API boards such as the UP2000 dual-processor motherboard or the single-processor UP1000, drop them into standard rack-mount enclosures, and populate them with memory, disks and network adapters sourced on the open market. The processor itself, detailed on the 21264 processor page, is the same part used in far more expensive systems.
What customers gain
- Price/performance: clusters built from commodity-form-factor Alpha nodes cost substantially less per unit of sustained floating-point performance than traditional supercomputers of the era.
- Scalability: a system can start with a handful of nodes and grow to dozens or hundreds as budgets allow, without replacing the original investment.
- Open software: Linux, the GNU toolchain, MPI and PVM message-passing libraries, and Compaq's optimized Alpha compilers and math libraries all run on the nodes.
- Serviceability: a failed node can be pulled and swapped while the rest of the cluster keeps working, and spare parts are standard components.
How a Linux NetworX Alpha cluster is built
A typical configuration pairs a master or head node, which handles logins, job scheduling and file serving, with a large number of compute nodes. Each compute node is an API Alpha board with one or two 21264 processors, its own memory, and at least one network interface. Fast Ethernet is sufficient for loosely coupled workloads, while applications that exchange data frequently benefit from a low-latency interconnect such as Myrinet or SCI. Linux NetworX adds its own management layer so that administrators can install, monitor and power-cycle nodes from a single console rather than visiting every machine.
| Component | Typical choice (2000) | Notes |
|---|---|---|
| Compute node board | API UP2000 or UP1000 | One or two Alpha 21264 processors per node |
| Processor clock | Roughly 600-750 MHz | Depends on board and module generation |
| Backside cache | Several MB per CPU | Dedicated off-chip cache for high hit rates |
| Memory | Up to several GB per node | ECC SDRAM |
| Interconnect | Fast Ethernet, Myrinet or SCI | Chosen to match application communication patterns |
| Operating system | Linux for Alpha | Red Hat, SuSE or similar distributions |
| Parallel libraries | MPI, PVM | Standard message-passing interfaces |
Who the systems are for
The early market for Alpha Linux clusters was concentrated in places where floating-point performance was the whole point. National laboratories and university research groups ran physics, climate and chemistry simulations. Oil and gas companies used clusters for seismic imaging. Aerospace and automotive engineers ran crash and airflow simulations. Life-science groups were beginning to use clusters for sequence analysis and protein modeling. In each case, the buyer was comparing cost per sustained gigaflop, and a well-built Alpha cluster compared favorably with both proprietary supercomputers and clusters made from 32-bit x86 nodes of the same period.
Cluster builders want a node that is fast, stable and easy to integrate. Our job is to make the Alpha the easiest high-performance node to build with, so partners like Linux NetworX can concentrate on the system and the customer.
An API NetWorks spokesperson
Part of a broader partner strategy
The Linux NetworX relationship is one of several ways API has worked to move Alpha from a niche workstation part into a commodity building block. Rather than building and selling finished systems in competition with its customers, API supplies boards, processors and reference designs to integrators, OEMs and value-added resellers, many of which are listed on the partners page. Cluster integrators are an especially good fit for that model: they need a steady supply of identical, well-documented nodes, and they add value through integration, software and support rather than through custom hardware.
The same approach runs through other API announcements of 2000, including the Dolphin and Scali WulfKit for Alpha Linux bundle, which pairs a low-latency interconnect and cluster software with API boards, and the later move toward very dense, rack-optimized servers designed with clusters in mind.
Availability
Linux NetworX offers Alpha-based cluster configurations built on API technology directly to end customers, with configuration, integration, burn-in testing and support provided by Linux NetworX. API boards and processors used in these systems are available through API's normal channels, and technical documentation is available through API support. Pricing depends on the number of nodes, processors per node, memory and interconnect selected.
About API NetWorks, Inc.
API NetWorks, Inc., formerly Alpha Processor, Inc., is headquartered in Concord, Massachusetts. The company was founded in 1998 as a joint effort of Samsung and Digital Equipment Corporation, later Compaq, to bring the 64-bit Alpha RISC architecture into mainstream markets. API designs and markets Alpha motherboards, processor modules, high-density rack servers and system-level building blocks for Linux and Tru64 UNIX, and supplies them to OEMs, system integrators and cluster builders worldwide. Through its API NetWorks unit the company is extending its high-performance design expertise into interconnect silicon and platforms for networking and communications equipment.