So in my previous article we have discussed about the ‘glueless’ architecture. You may want to read part 1 before proceeding.
We have seen that ‘glueless’ architecture as some serious drawbacks. Let’s see if the second main scale-up server architecture can mitigate those issues. Meet…
The ‘glued’ architecture
We’ve seen that the ‘glueless’ architecture, coordination and communication between the processor sockets, creates a bottleneck. To overcome this problem hardware manufacturers have developed a ‘glue’ to the architecture. This ‘glued’ architecture uses external node-controllers to interconnect QPI island, that is kind of clusters of processor sockets.
Intel QPI links offers a scalable solutions based on OEM-developed eXternal Node-Controllers (referred to as XNC). External node-controllers using the Intel Xeon E7-4800 series with embedded memory controller implies a Cache Coherent Non-Uniform Memory Access (ccNUMA) system. The role of ccNUMA is to ensure cache coherency by tracking the most up to date data is for every cache line held in a processor cache.
Latency between processor and memory in a ccNUMA system varies depending on the location of these two components in relation to each other. Manufacturers also want to minimize the bandwidth consumption resulting from the coherency snoop (Intel QPI source broadcast snoopy).
Therefore the quality of the OEM-developed eXternal Node-Controllers is critical and only a few manufacturers are able to provide server architecture which scale in pace with resources added to the system.
The next article in this series, I will focus on Bull’s eXternal Node-Controller called the BCS. Stay tuned!
Source: Bull, Intel, Wikipedia