New Data Center Strategies Newsletter, By Andreas M. Antonopoulos, Network World, 12/12/06

Could RAM be moved off server blades?

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.

Thus, all the compute blades can draw from a single large pool. As with server virtualization, RAM virtualization would allow higher utilization of RAM and flexible allocation policies.

Here are some of the advantages:

* Higher utilization. Some blades may have more RAM than they need, some have less. Rather than re-allocating workloads or moving physical RAM DIMMS, RAM could be re-allocated virtually from a pool.

* Flexibility. As workloads change over time, applications could allocate as much or as little RAM as they need, regardless of the physical RAM present on the blade itself. With 64-bit operating systems, addressable RAM is much higher (16TB for Windows x64).

* Lower cost. Cost of RAM is driven by size (capacity) and density. Density is a problem for blades because they have only a limited number of DIMM slots. Virtualization would allow mixing 1GB, 4GB, 16GB and higher DIMM sizes so that RAM can be purchased based on density and cost, not based on limited slots.

* Component simplification. Just like moving storage off the server has made the server simpler - resulting in blades - moving RAM off would further simplify the design of the blade.

RAM blades are a highly speculative proposition, but they are in line with the developments we have seen in the data center. Obviously the development of RAM blades presupposes higher-speed networks in the data center, and that is also a clear trend. With RAM leaving the board, the server will be fully broken into individual resources, and virtualization will have achieved a plateau. These are indeed exciting times in computing and the data center.