HPC Cluster Blog – Cluster Interconnects

Cluster interconnects: the three basic configurations

While processor speed, memory amount and number of nodes in your cluster are important, another serious consideration for overall cluster performance is the interconnect. The term “interconnect” refers to the cabling and switches used by the nodes to communicate with each other. While necessary for every cluster, it becomes particularly important with highly parallel computing, where calculations are split among nodes to be solved simultaneously.

The three primary configurations we recommend to our customers are standard gigabit Ethernet, 10GbE (short for “10 gigabit Ethernet”) and InfiniBand. Each has its own pluses and minuses; which one to use depends on your needs, budget and setup.

For very small clusters or clusters in which fast inter-node communication isn’t critical, gigabit Ethernet, with transmission speeds of one gigabit per second, is perfectly acceptable. This form of interconnect uses standard Ethernet switches and Cat5e or Cat6 copper cable, the ubiquitous type with an RJ45 connector used in patch cables and standard LAN connections.

While providing slow communication speeds, gigabit is far less expensive than other interconnects and is easy to set up. We recommend gigabit to customers who are on an extremely tight budget and/or expect their nodes to be working independently most of the time.

However, the majority of our customers prefer a faster interconnect to facilitate inter-node communication and parallel jobs, which generally leads them to choose between either 10GbE or InfiniBand.

The 10GbE interconnect is an Ethernet connection that operates at 10 gigabits per second, 10 times faster than standard gigabit Ethernet with less latency (delays in data transmission). 10GbE can take advantage of either standard copper cabling for short distances or optic fiber for longer ones. A 10GbE interconnect has typically been more expensive than the faster InfiniBand setup in both hardware and power requirements, but the newest advances in 10GbE technology are reducing these disadvantages. The strongest advantage for 10GbE is that it uses standard Ethernet protocols, making it easier for Ethernet-trained technicians to adopt and use, and is backward-compatible with existing gigabit networks.

InfiniBand, currently our most commonly sold interconnect, operates with a different set of protocols than the Ethernet-based interconnects mentioned above. Due to its uniqueness, cluster administrators need specialized training or self-education to use this type of communications link. However, InfiniBand outpaces 10GbE with both its low latency and fast speeds, the most common at this time being either 40 or 56 gigabits per second. Other advantages over 10GbE is that it has been up to this point cheaper, less power-hungry and readily available as an on-board port on many server motherboards.

Unlike Ethernet’s switched network, which allows switches to be daisy-chained together to add ports, InfiniBand uses a fabric topology, meaning that every node is connected to every other node. In large clusters, this often results in multiple switches with a plethora of cables connecting each one to the other. While this may sound extreme, this “fat tree” configuration insures InfiniBand’s high throughput rate, availability and scalability.

If you’re unsure of which interconnect is right for your cluster, please contact us at [email protected] to discuss these options. We can help you weigh the pros and cons and determine the right solution for you.