#
- OSs not aware of difference between local or NUMA memory access
- Resource management policies not setup for large number of processors
- Use virtual machine monitor to run unmodified OSs on scalable multiprocessors
- VM monitor takes care of NUMA awareness
- HW failures confined to a single cell of Cellular Disco
- Overcommittment of physical resources is fine and even encouraged
- Cellular disco assigns a fraction of these resources to each virtual machine depending on their needs
- Multiplexes physical processors among virtual machines
- Cells intercept privileged operations by VMs
- Virtual machine code runs in ring 1 - supervisor mode
- Cellular disco code runs in ring 0 - kernel mode
- pmap data structure maps virtual and physical addreses
- TLB flushes are avoided because Cellular Disco mantains a Software TLB per vCPU (L2TLB)
- A 1000 core processor cannot affor a reboot when 1 processor fails, it's too common of an event
- Only applications running in faulty processors will be lost
- Structure of multiple cells who contain complete copy of monitor text and who manage pages belonging to its nodes allow for independent failures
- Cells also share memory for IPC
- Each machine processor maintains a run queue of vCPUs
- Load balancing via vCPU migration from one processor to another is good in case of overloaded processors
- Each cell manages its own memory allocation
- Memory borrowing from other cells is OK when a cell is memory starved and others have free memory
- Large applications can bypass the OS and claim large shared memory regions by directly interacting with the virtual machine monitor
- Cellular disco needs to make OS not page out these memory regions
- Dominant cost of overhead = kernel traps
