|
|
HPC InterconnectsOverviewA high performance computational cluster is only as fast as the weakest link. And if that link is the interconnect technology between cluster nodes, you could very well find that you have just deployed the worlds fastest door-stop! The ideal HPC cluster configuration has a balanced specification of processing power, memory capacity & bandwidth and fast inter-node communication channels. There are two common ways to improve inter-node communication bandwidth:
Channel bonding involves combining two or more slower communication channels to build-up performance by load-balancing the data flow. Typically the communication channels consist of Fast or Gigabit Ethernet ports integrated on the system motherboard. This method is extremely cheap as it utilizes ubiquitous Ethernet technology and requires very little additional expertise to configure, deploy and manage. The disadvantage with this method is that while net bandwidth may increase, so would latency. Also some applications may need to be altered to take advantage of the load-balancing nature of a bonded communication channel. Over the years, as multi-node HPC clusters have taken a precedence over traditional vertical scaling supercomputing technology, a new breed of interconnect options have been developed. These are highly specialized ASICs and adapters that are tuned specifically for the high-bandwidth, low-latency requirements of HPC cluster environments and applications. The idea is that interconnect technology should be able to move large amounts of data quickly enough around 10s or even 100s of nodes, that the cost-performance-scalability ratio of HPC clusters significantly overshadows that of vertical-scaling SMPs. Highly specialized or standards-based high-speed interconnects augment in-built Gigabit Ethernet technology. If your application message passing bandwidth and latency requirements cannot be met by gigabit Ethernet, the following high-speed interconnects provide excellent alternatives. Host Channel Adapters
|