Bare-Metal Container: Direct Execution of a Container Image on a Remote Machine with an Optimized Kernel

Container technology has become popular in HPC applications because it offers easy customization and quick execution. However, container technology does not allow kernel-level optimization. It does not allow changing a kernel as well as loading/unloading a kernel module, thus it prevents the extraction of full performance on a target machine. In addition, container technology is designed for multi-tenancy and makes difficult to measure exact performance and power consumption for an application on a target machine. Bare-Metal Container (BMC) solves these problems by a mechanism that offers a suitable kernel for a container image and boots it on a remote physical machine. The application can occupy the physical machine and utilize the full performance. It is implemented by remote machine management technologies (WakeupOnLAN, Intel AMT, and IPMI), and a network bootloader. The power consumption is also measured, and the user knows the good combination among application, kernel and target machine. Although BMC requires the overhead of kernel booting, HPC applications can improve performance, which surpasses the overhead. The experiments showed the effect of kernel optimizations for CPU (Hyper Threading), memory (Transparent Huge Pages), and network (Receive Flow Steering) from a low power CPU (Celeron) to a high speed CPU (Xeon).