VM Placement in Accelerator-Equipped Data Centers Using Variable-Length Modified Genetic Algorithm

With the increasing growth in computing demand and the complexity of applications, cloud computing has become very popular in recent years. To cope with the high demand for computing resources, data center providers have introduced various types of hardware accelerators such as GPUs and FPGAs in their environments. Accelerator virtualization is also introduced to overcome the underutilization of GPUs and FPGAs in such environments. However, resource provisioning can be quite challenging in large data centers with heterogeneous setups due to the massive solution space of the resulting VM placement problem. In scenarios where the VM resource requests exceed the amount of available resources on the hosts, finding an energy-efficient solution with maximum VM acceptance rate can get difficult for simple heuristics or even metaheuristic methods under tight decision time constraints. The aim of this paper is to tackle the inefficiency of genetic algorithm (GA) in producing good partial schedules in limited decision time. To this end, we introduce a GA-based VM placement method called VLMGA (variable-length modified genetic algorithm). Starting from a limited solution space, VLMGA iteratively tries to find a solution in each sub-space and enlarge the search space until no feasible solution could be found within the specified time frame. Using the proposed technique, the quality of the obtained solution can be greatly improved. Evaluated under real-world workload scenarios, the proposed method achieved 16% improvement on the energy-delay product compared to well-known VM placement methods.