ColocationSim: Simulate Colocation Datacenter with Microservices and Performance Interference
Kangjin Wang, Ying Li · 2021 IEEE International Symposium on Software Reliability Engineering Workshops (ISSREW) · 2021
Colocating Latency-Critical (LC) jobs and Best-Effort (BE) jobs on one host is the best practice to improve datacenter resource efficiency and the Colocated Job Scheduling algorithms play a key role in this scenario. But validating a new Colocated Job Scheduling algorithm at scale is usually resource-, energy- and time-consuming in the real-datacenter, therefore simulation is the first choice. Unfortunately, existing simulators are not suitable for colocation simulation because: 1) they cannot simulate microservices, which is widely adopted in LC/BE jobs; 2) they cannot simulate performance interference of colocated jobs. In this paper, we present a new simulator, ColocationSim, which 1) uses REG(Request Execution Graph) model and queue theory to simulate the behavior of microservices; 2) uses pre-trained models to simulate performance interference. Experiments show that ColocationSim can simulate performance interference of both sequential and fan-out structure microservices with high accuracy and it can also simulate more than 8,000 colocated hosts in a short time.