Disaggregated and optically interconnected memory: when will it be cost effective?

Abstract

The "Disaggregated Server" concept has been proposed for datacenters where the same type server resources are aggregated in their respective pools, for example a compute pool, memory pool, network pool, and a storage pool. Each server is constructed dynamically by allocating the right amount of resources from these pools according to the workload's requirements. Modularity, higher packaging and cooling efficiencies, and higher resource utilization are among the suggested benefits. With the emergence of very large datacenters, "clouds" containing tens of thousands of servers, datacenter efficiency has become an important topic. Few computer chip and systems vendors are working on and making frequent announcements on silicon photonics and disaggregated memory systems. In this paper we study the trade-off between cost and performance of building a disaggregated memory system where DRAM modules in the datacenter are pooled, for example in memory-only chassis and racks. The compute pool and the memory pool are interconnected by an optical interconnect to overcome the distance and bandwidth issues of electrical fabrics. We construct a simple cost model that includes the cost of latency, cost of bandwidth and the savings expected from a disaggregated memory system. We then identify the level at which a disaggregated memory system becomes cost competitive with a traditional direct attached memory system. Our analysis shows that a rack-scale disaggregated memory system will have a non-trivial performance penalty, and at the datacenter scale the penalty is impractically high, and the optical interconnect costs are at least a factor of 10 more expensive than where they should be when compared to the traditional direct attached memory systems.