Kernels

the type of kernel that you use, and the OS architecture, depends on the application

Monolithic kernels

• main program invokes syscall
• kernel is underlying monolithic block:
  • service procedures carry out syscalls
  • utility procedures help implement service procedures

• separate applications and OS using privilege levels into user and kernel
  • on x86, 4 privilege levels (but in practice mostly 2 are used)
  • this is supported by the hardware directly
  • if only goal is to separate untrustworthy apps from lower level shit, you only need 2 separate levels
  • if you include more levels, there’s a cost associated with switching between levels, so why do it if it’s not needed

Virtualisation

• originally to separate multiprogramming from extended machine
• N independent system call interfaces

• Virtual machine monitor (VMM/Hypervisor) emulates hardware
• types:

  • 1: VMM runs on bare metal (like Xen)

    

  • 2: VMM hosted on OS (like QEMU)

    

  • Hybrid: VMM inside OS (like KVM)

Exokernel

• separate resource control from extended machine
• unlike VMM/Hypervisor, it:
  • does not emulate hardware. only resource manager
  • only provides safe low-level resource sharing
  • service procedures are offered as library linked directly to application – “Library OS”
• different library OSes for different programs, allows application-level specialisation

Client/server model (microkernel)

• organise service procedures in programs running in separate processes (system services/drivers)
• high level of isolation
• processes communicate via message passing
• calls rely on the same mechanism (message passing)
• messaging is implemented in microkernel (minimal kernel)
• principle of least privilege – isolate every service in its own domain (address space, process, etc.)
• this is more secure, but lower performance (always a tradeoff). have to switch between modes and shit

Microvisor

• combination of hypervisor and microkernel
• different OS architectures have different design points, people look at convergence and tradeoffs

Unikernel

• “single simple application implementing whatever in the cloud, in most efficient way”
• squash application and OS kernel into one thing, don’t need all of the other stuff like process management
• gets rid of all of the overhead