Features:

DELL'S ROOHOLAMINI ON THE STATE OF LINUX CLUSTERS
by Derrick Harris, Assistant Editor

HPCwire interviews Reza Rooholamini, Dell's Director of Operating Systems and Clustering about the current state and future of Linux clustering. He is a keynote at the upcoming Linux Clusters Institute's Linux HPC Revolution 2004 Conference.

HPCwire: What are the factors that currently contribute to growth of the Linux HPC market, and in particular what are the major inhibitors that challenge that growth, both technically and economically?

REZA ROOHOLAMINI: Many factors contribute to this growth, but the two dominant factors are: 1) the availability of high-performing hardware building blocks based on standardized technology, and 2) the emergence of Linux as an enterprise operating system.

Intel processors have greatly increased in performance over the past decade, while dramatically decreasing in price. Similarly, Linux has matured as an operating system, with better SMP support, networking stacks, and security, among other features, while maintaining its tradition of openness and public accessibility. We believe that the combination of the two has led to the growth that we see today in Linux-based HPC clusters.

We do not see any major inhibitors in this growth, but recognize the importance of a few important enablers: With larger deployments the installation, management, configuration, and administration of clusters have become more complex. To reduce this complexity, customers need tools that automate routine cluster administration tasks. There needs to be more research and development in this area and Dell has invested its engineering resources to simplify the deployment and management of HPC clusters. We feel that this is critical to the continued acceptance of clustering as a cost-effective paradigm for supercomputing. Maintaining attractive price/performance is an economic imperative and clearly requires careful selection of support models and partners according to changing global realities.

HPCwire: To what extent is current Linux cluster deployment influencing future growth?

RR: Current Linux cluster deployment is paving the way for future growth. In general, customers deploy Linux clusters because they offer an attractive price/performance option. This is unlikely to change. We see a trend where customers start by deploying a small cluster and as their needs grow, they add more hardware to increase the size of their cluster. This "pay as you grow" model allows customers to buy only the needed computing power today, and maximize their return on investment. Additionally, clusters are heterogeneous entities by definition, which allows for many generations of hardware and software to work together on solving the same problems This inter-operability, which is enabled by following standards, allows customers to protect their IT investments by prolonging the life of existing assets and integrating the latest technology as they become available.

HPCwire: What changes are in store for Linux HPC clusters in the next 5 years?

RR: We foresee a few interesting trends emerging: 1) We have seen our enterprise corporate customers increasingly adopt clustering, which will lead to improvements in reliability, availability and serviceability of business systems as clusters proliferate. 2) We see many standardization efforts on the way in the software and hardware arenas. For instance, Infiniband is replacing the proprietary interconnects typically used in this space and will provide a better value to our customers. Similarly, open source initiatives like OSCAR and ROCKS will continue to evolve the software management needed for HPC clusters, lowering the barrier to entry for first-time cluster builders. 3) We see larger clusters (in excess of 1,024 nodes) becoming commonplace. If you look at past few Top500 supercomputing lists, you can clearly see an upward trend in the number of processors per cluster. 4) As organizations own multiple clusters, customers will begin to connect them into grids to help them use their IT assets more efficiently.

HPCwire: Who will be the main users 5 years from now?

RR: We believe everyone will use clustering technology in the coming years, whether directly or indirectly. As it is today, the use of clusters to solve engineering and scientific problems will continue, and we will see increasing numbers of petroleum engineers, chemists, biologists, and financials analysts using clusters to solve their respective problems. Clusters will eventually become transparent to the organizations. For example, in the future customers will perform online searches, analyze their stock portfolios and check the weather via the Internet from applications powered by clusters that are geographically dispersed. The users will not have knowledge of where and how the information is located and processed. As this technology becomes pervasive greater standards will emerge, further driving down the cost and increasing simplicity.

HPCwire: Are users leading the path or are they merely adapting to the technology path that is presented to them?

RR: Our users are driving this revolution by giving us requirements for faster, lower latency interconnects, more scalable file systems, more manageable clusters, more secure operating systems, and grid-enabled clusters, among other requirements. Many of these technologies have their origin in our customer space, particularly in academia and governmental laboratories. Two examples: 1) ROCKS is driven by a consortium called NPACI, led by San Diego Supercomputing Center, and other prestigious institutes across the country. 2) Lustre is a scalable file system, envisioned by the tri-labs in the US, and now being commercialized by Cluster File System. We see ourselves as integrators and packagers of these technologies. Dell is investing our R&D to develop useful solutions for our customers.

HPCwire: Has there been any real impact of open source to the growth of the Linux HPC market, and if so please cite specific examples?

RR: Linux has a well developed, mature and vibrant eco-system around it. Numerous open source efforts have aided the growth of HPC clusters. In the area of file systems, the Parallel Virtual File System project, which started at Clemson University, and later adopted by Argonne National Lab, has provided a scalable file system for Linux-based HPC cluster users. In the area of cluster administration, Ganglia is an open-source distributed monitoring system for clusters from University of California, Berkeley. These products have proven very useful to users and have had a positive impact on the growth of Linux clusters. The examples of ROCKS and OSCAR given earlier also complement this list.

HPCwire: Is there any future to "value-add" Linux systems, or will economics eventually squash all attempts to increase vendor profit margins?

RR: The key here is to maintain the benefit of the "value-add" components to the customer. Standardization maximizes this benefit and offers Dell an opportunity to provide a spectrum of solutions at approximately one-tenth the cost of a proprietary supercomputer. We design solutions with clear progression paths, to maximize adaptability without compromising growth potential. For instance, Dell HPC clusters offer design choices to customers at different levels: A cluster can be built by using standard based Ethernet technology or by using faster Quadrics, Myrinet, or SCI fabrics. The same multitude of choices is true for file systems, Message Passing Interfaces, compilers, and application libraries.

Dell delivers value and ensures high performance for customers by testing, integrating, benchmarking and qualifying these technologies, and packaging them into a turn-key solution. Another value-add is providing deployment services for large clusters, where pre-sales consulting, installation, configuration, and optimization greatly simplifies deployment. We believe that by giving our customers the services and products that they require, they will continue to partner with us for their high-performance computing needs.

HPCwire: How do businesses plan to cope with this pressure to reduce cost by eliminating these value-add proposals?

RR: The pressure to drive down cost is present across the computing industry. This trend underlines the importance of scalable standards-based solutions, such as those Dell offers, that are high-performing, flexible and cost- effective -- therefore, more attractive to organizations. A "value-add" by definition should justify its cost. Partners must give customers the choice to select the right "value-add" components for their needs today, without locking them in. It is also important that these solutions are flexible enough grow with technological innovations that address their business goals tomorrow.

Dell has the flexibility to build clusters with open source packages, or commercially-sourced ones. These choices offer a range in performance, functionality, and price. Customers can grow their clusters, add new technology, and acquire additional services as their needs grow. Dell's approach to providing industry-standard HPC solutions is a reflection of the Dell direct model, where we focus our R&D efforts on solutions that customers[are demanding, thus maximizing resources based on volume and driving down the cost of the total solution. It is a model that has garnered success in the HPC market today.