HPC Cluster Solutions by Penguin
Besides latest x86 servers running Linux, Penguin Computing provides also complete cluster solutions for High Performance Computing and rack level integration and delivery of scale-out server farm environments.
True integrated cluster is more than just a collection of computers. Cluster is essentially a large distributed computer consisting of multiple nodes connected together. Complexity, physical size and power requirements pose their own challenges. Penguin has solutions.
Both HPC clusters and rack level builds leverage the fine tuned, time tested, ISO 9000 compliant integration, burn-in and testing process at the Penguin factory. Our mission is to produce robust, consistent hardware with components selected and preconfigured to run the commercial and free versions of the Linux operating system.
For more information regarding specific subtopics and Penguin’s preferred hardware partners please see below.
Penguin also offers a complete range of professional services and expertise to help simplify the planning, design and implementation of your cluster system or datacenter environment.
Cluster Interconnects
InfiniBandInfiniBand is a high bandwidth, low latency interconnect technology for reliable, scalable, high performance communication fabrics. Current InfiniBand technology generation is called Quad Data Rate (QDR) and provides 40Gbps host and switch-to-switch links.
Penguin offers InfiniBand switches and host adapters from QLogic and Mellanox. Host side software stack is available with commercial and free Linux operating system distributions.
10 Gigabit Ethernet10 Gigabit Ethernet (10GigE) provides 10x boost in bandwidth compared to ubiquitous Gigabit Ethernet and inherent interoperability with existing Ethernet networks. Use cases for 10GigE include congestion elimination on busy switch uplinks, storage server attachment and low latency messaging networks for high frequency trading.
Penguin Computing sells fixed configuration and modular Gigabit and 10 Gigabit Ethernet switches from leading manufacturers listed below.
For 10 Gigabit Ethernet host adapters supported under Linux operating system Penguin recommends
Cluster Deployment and Management
Due to the number of hardware units in typical HPC clusters and large server farms initial deployment, practical day-to-day administration and preventative maintenance requires automated approach.
Penguin’s Scyld ClusterWare supports automated deployment of large diskless compute clusters, ensuring consistent software versioning and configuration. This technology was pioneered by Donald Becker, longtime CTO of Penguin Computing.
All Penguin server hardware includes integrated Baseboard Management Controller (BMC) which implements the remote monitoring and management on node level. For rack and cluster level remote control Penguin recommends KVM (Keyboard-Video-Mouse) and KVM-over-IP appliances from our partners:
Power Delivery and Cooling
High end multi-socket, multi-core HPC systems can use close to 1000 Watts per rack unit, especially when equipped with GPUs. Common equipment rack stuffed with these systems and some room for networking gear can approach 30 to 40 kilowatts. This level of current per circuit and power density per square foot requires specialized power delivery and heat eviction equipment.
At the other end of the power spectrum are the customers with large web server farms over provisioned to handle the daily or periodical usage spikes. Rather than a compute bound simulation job large scale web applications are typically network and I/O bound, resulting in continuous operation at partial utilization level. In these environments Total Cost of Ownership can be greatly enhanced by optimized airflow designs, careful matching of the power supplies to the load and by selecting the system components with superior power characteristics.
Penguin Computing delivers value for both high power and high efficiency environments by pulling from 12+ years of system and solution design expertise and strong line card of datacenter partners.
