dReDBox: A Disaggregated Architectural Perspective for Data Centers

Abstract

Data centers are currently constructed with fixed blocks (blades); the hard boundaries of this approach lead to suboptimal utilization of resources and increased energy requirements. The dReDBox (disaggregated Recursive Datacenter in a Box) project addresses the problem of fixed resource proportionality in next-generation, low-power data centers by proposing a paradigm shift toward finer resource allocation granularity, where the unit is the function block rather than the mainboard tray. This introduces various challenges at the system design level, requiring elastic hardware architectures, efficient software support and management, and programmable interconnect. Memory and hardware accelerators can be dynamically assigned to processing units to boost application performance, while high-speed, low-latency electrical and optical interconnect is a prerequisite for realizing the concept of data center disaggregation. This chapter presents the dReDBox hardware architecture and discusses design aspects of the software infrastructure for resource allocation and management. Furthermore, initial simulation and evaluation results for accessing remote, disaggregated memory are presented, employing benchmarks from the Splash-3 and the CloudSuite benchmark suites.

Design, Modeling, and Analysis of Memory Allocation Policies for Rack-Scale Memory Disaggregation

“…We felt that approximation of the memory request cycle is good enough to quickly study the performance of large-scale systems where memory is the main concern, for which we integrate a cycle-accurate memory simulator. We further modify the base tool to support multi-threaded workloads, over which instrumentation is performed at instruction-level granularity to collect LLC misses, as done by [31]. A single trace is produced by merging the cache misses of all the cores sorted by their time stamps.…”

“…For every page fault, it allocates a new page in either local or remote memory based on the allocation policy. All the memory accesses get divided among local and remote memory addresses as per their page location in the address space, rather than using a static approach as in [2,20,31]. The global memory manager is also simulated to serve the remote memory requests from the compute nodes.…”

Empowering Data Centers for Next Generation Trusted Computing

“…Setting & Problem Statement. Modern data centers enable disaggregated execution (e.g., dReDBox project [39,40]) to improve the performance of computation by using domainspecific accelerators (DSA) such as GPUs, ML/AI accelerators, storage accelerators [41], and smart NICs [42]. Fig.…”