On Tue, 19 May 2020 08:10:00 +0930 Brett Lymn <bl...@internode.on.net> wrote:
> On Sat, May 16, 2020 at 09:51:42AM +0100, Sad Clouds wrote: > > > > Just look at how Solaris does it - it has Zones (aka Jails) and > > LDOMs (Logical Domains) on SPARC. LDOMs seem to be a much better > > way of partitioning OS instances versus something like VMware or > > Xen. > > > almost but not quite. A SPARC LDOM is more of a hardware > partitionig, cpu and memory are dedicated to the LDOM for its > exclusive use. You can configure some dynamism by allowing cpu to be > added/removed automatically when the load average increases/decreases > but this relies on the guest doing dynamic reconfiguration. That's exactly what I was referring to. Yes this is specific to SPARC where they have a very small firmware hypervisor. The advantage is how hardware resources are dedicated to a specific domain, so the OS can use them directly with very little overhead. Xen supports "PCI passthrough", which can reduce the overheads when a guest accesses that device. But I don't think it has anything like "CPU or memory passthrough", it supports "CPU pinning" however I think you still end up with excessive cache and TLBs thrashing. I haven't actually done any detailed benchmarking at the hardware level, so could be talking complete nonsense. This is just the impression I got when reading various technical manuals over the last few years.
