Take one simple example: virtualization versus ‘bare metal’ performance. Many assume that virtualized workloads are always slower than those running on ‘bare metal’. But what those people fail to realize is that the application is not running directly on ‘bare metal’, but indeed on top of an OS. And this operating system can also introduce its own overheads. One particularly interesting example of this is with memory-bound workloads, such as big data or AI type workloads. What we’ve found is that running these workloads on VMware vSphere actually outperforms Linux ‘bare metal’ systems by around 5% in some scenarios. How can this be, given that vSphere has some virtualization overhead (even as minimal as it is these days)? Well it turns out that VMware ESXi has far superior NUMA scheduling capabilities than Linux and because of that memory access latencies are lower with VMware ESXi, giving it a slight performance advantage. (Of course, this book delves quite deeply into NUMA architecture!
