Over the last five years, servers have been deconstructed into different parts, and some of those parts have ended up on a network. Servers have transformed into blades, and power supplies, network interfaces and storage are shared across many blades, which contain just the basic CPU and memory.

In many organizations the majority of storage systems are SAN-based, shared over a high-speed, fiber-optic network. Many other services have also migrated to networks, such as SSL termination and acceleration, security services (firewalls and IDS/IPS) and optimization services (TCP acceleration, caching and protocol optimization).

The “minimalist” architecture of blades has evolved out of faster networks. As data center networks get faster, could the next step be splitting RAM from the blade servers?

At the moment, rapid access to RAM requires a high-speed bus with very low latency, so RAM is located on the same board as the CPU and is available only to a single blade. It is conceivable that RAM could end up splitting into two parts: a smaller RAM chip on the blade for caching and a bigger pool of RAM in a separate blade or even served across a high-speed network.

In this type of architecture, RAM could be partitioned and allocated to different blade servers on demand, with virtual machines able to expand and contract their RAM needs dynamically. Of course, even a high-speed network would be slower than an on-board bus, but it is possible that the additional flexibility would be a greater benefit than the performance impact.

RAM blades would look like a compute blade but would just include rows and rows of DIMM slots. A single RAM blade could provide dozens or even hundreds of gigabytes of RAM. Compute blades would then be simplified, with just a CPU and perhaps a small amount of RAM for caching.

In essence, there would be three levels of memory cache: L1 and L2 cache on the CPU, and an additional small RAM chip on the blade (L3?). In today’s blades, the RAM is slotted into DIMM slots and is upgradeable. In a compute blade, the on-board RAM might be fixed. The majority of the RAM needs of a compute blade would be provided by allocation of RAM from a pool of RAM blades, either in the same blade frame, or in an adjacent rack.