Abstract <p> A new dynamic bin packing problem relevant to cloud computing is considered. Thecreation time, deletion time, and required resources are known for each item (virtual machine).The containers (servers) have a NUMA architecture and specific rules for placing the machines.The servers are grouped into racks, and some machines form groups. Each group is divided intopartitions. Machines from different partitions cannot be placed on the same rack to ensure systemfault tolerance. The objective is to pack all the machines into the minimum number of racks overa given planning horizon. A two-stage algorithm is developed to solve the problem: an initialsolution is constructed, where some constraints may be violated, followed by iterativeimprovement using local search aimed at eliminating the violations. Using the proposed approach,an average deviation of 3.8% from the lower bound was achieved on open test cases.</p>