From Nominal Computing Capacity to Effective Service Capacity: Resource Complementarity, GPU Compatibility, and Investment Allocation
Year: 2026
Type: preprint
Abstract: Idle computing capacity can coexist with unserved tasks, making investment efficiency depend on compatible resource combinations and allocation rules. We develop an integrated decision framework for combining hardware investment and allocation management to expand effective service capacity. A mixed-integer model links task admission to node-level resources, graphics processing unit (GPU) compatibility, and placement rules. Computational experiments use Alibaba GPU v2023 data. Under common initial conditions, central processing unit expansion admits 99 tasks versus 97 for GPU expansion without model restrictions, but 77 versus 81 with them. Compatibility thus reverses the service-quantity ranking of the two investments. With 20 additional GPUs and no extra per-node installation cap, targeted investment raises admission from 81 under uniform expansion to 95; pooling and repacking on that hardware reach 98, matching an independent upper bound. Despite overlapping stand-alone gains (a quantity interaction of -7), the combined strategy enables service targets of 96-98 tasks that neither intervention can achieve independently within the specified project menu. Installation constraints and heterogeneous task values further condition investment choices. Verified allocations, independent upper bounds, and cost thresholds translate resource constraints into investment decisions. Compatible resource bundles and allocation flexibility determine effective service capacity and the service returns to computing infrastructure investment. MSC Classification (2020): 90C10 , 90B50 ORCID (Jie Liu): 0009-0007-2422-2280
Source: Research Square
Author: Jie Liu
Institution: Guangzhou College of Commerce
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225987241
