Slide 2 of 24
What we have on the market for clusters are Low End and High End designs for processing nodes.
Up to this date, high end systems and low end systems still differ significantly:
Low end systems feature extremely low cost, but much weaker memory systems.
In addition all I/O operations (including Gigabit networking) must be performed over a standard I/O interface, the PCI-Bus, which delivers lower bandwidth than the host and the memory bus.
High end systems on the other hand include carefully engineered memory systems. They include special hooks for any additional datastream caused by memory intensive computations or by inter-node communication.
That raises the following question about parallel and distributed compute nodes.
Are Low End SMPs the universal compute nodes for parallel and distributed systems?