Data Center Resource Disaggregation¶
|Mar 7, 2022
This note discusses Resource Disaggregation’s Design Spectrum.
In our categorization, the traditional distributed systems approach is logical resource disaggregation. The newly emerged hardware resource disaggregation is physical resource disaggregation. The main difference lies in whether an indirection layer is required to achieve the conceptual resource pool view. Combined, they are two extreme design points of the resource disaggregation idea.
All images below are from my recent defense slides. This note is part of my defense’s intro.
The full defense slide is here.
Resource Disaggregation is a really general idea with a wide design spectrum that covers many designs and systems in data centers. The essense of resource disaggregation is to decouple resources so as to achieve independent resource scaling and failing. It has been applied in different granularities and to many different domains.
The traditional resource disaggregation is usually built on top of monolithic servers using conventional distributed systems. It has been applied everywhere in data centers, just in different granularities. For example, in the classical storage disaggregation deployment, storage pools are disaggregated from compute pools; in machine learning deployment, paramemter servers are disaggregated from workers; in typical SDN deployment, control plane servers are disaggregated from data plane servers/switches. All these examples are instantiations of the resource disaggregation idea.
Hardware Resource Disaggregation is a super HOT research proposal that breaks the physical monolithic servers into segregated, network-attached hardware resource pools, each of which can be built, managed, and scaled independently. The disaggregated approach largely increases the management flexibility of a data center.
Hardware resource disaggregation is a drastic depature from the traditional computing paradigm and it calls for a top-down redesign on hardware, system software, networking, and applications.
Is hardware resource disaggregation just a buzzword? Is it just another old wine in the new bottle kind of idea?
I argue that the traditional resource disaggregation design approach using distributed systems and the newly emerged hardware resource disaggregation are not exclusive to each other and in fact can be unified within one design spectrum, with each being one end of the spectrum.
Before we dig into the design spectrum. I want to spent a few words on the Resource Disaggregation Formula: one would take a set of system software and a set of disaggregated hardware devices or servers, then use whatever approach, to produce the same ultimate goal, which is the conceptual resource pool view. The pool can be a CPU pool, a memory pool, a Parameter Server pool. Basically every standalone “conceptual” resource. Think about the examples we mentioned earlier, all systems follow this formula, just produce different “resource pools”.
Now, the categorization.
On the far left, we have the logical resource disaggregation, which represents the traditional resource disaggregation model. This model builds on top of monolithic servers. A server would contribute part or all its resource to a certain resource pool. A server can be a part of multiple pools. Usually, an indirection layer at each server is required to achieve this goal. Essentially, the ultimate resource pool just logically maps back to the actual servers. This is the common-wisdon on building distributed systems.
On the far right, we have the physical resource disaggregation, which represents the emerging hardware resource disaggregation model. This model builds on top of disaggregated hardware devices. Usually, no indirection layer is required. So essentially, the ultimate resource pool could physically maps back to the actual physical devices.
In the middle, we have the Hybrid Disaggregation which has the best of both worlds. It has both normal servers and disaggregated devices.
The following image shows the design spectrum.
So far, my work in this space has covered all grounds. (DUH! I defined the specturm to fit my work! :-) )