Slashdot Mirror
Linux Kernel 4.10 Officially Released With Virtual GPU Support (softpedia.com)
"Linus Torvalds announced today the general availability of the Linux 4.10 kernel series, which add a great number of improvements, new security features, and support for the newest hardware components," writes Softpedia. prisoninmate quotes their report:
Linux kernel 4.10 has been in development for the past seven weeks, during which it received a total of seven Release Candidate snapshots that implemented all the changes that you'll soon be able to enjoy on your favorite Linux-based operating system... Prominent new features include virtual GPU (Graphics Processing Unit) support, new "perf c2c" tool that can be used for analysis of cacheline contention on NUMA systems, support for the L2/L3 caches of Intel processors (Intel Cache Allocation Technology), eBPF hooks for cgroups, hybrid block polling, and better writeback management. A new "perf sched timehist" feature has been added in Linux kernel 4.10 to provide detailed history of task scheduling, and there's experimental writeback cache and FAILFAST support for MD RAID5... Ubuntu 17.04 (Zesty Zapus) could be the first stable OS to ship with Linux 4.10.
It required 13,000 commits, plus over 1,200 merges, Linus wrote in the announcement, adding "On the whole, 4.10 didn't end up as small as it initially looked."
It required 13,000 commits, plus over 1,200 merges, Linus wrote in the announcement, adding "On the whole, 4.10 didn't end up as small as it initially looked."
My understanding is that it is more extensive: PCI(mostly 'e' these days) passthrough allows you to assign a physical device to a VM; but the device can't be shared: if a given piece of hardware is being passed through to one of the guests, none of the other guests or the host OS can use it.
This 'virtual GPU' stuff is supposed to make allocating GPU resources between VMs closer to how it is with CPU time or memory, where all the guests and the host can't exceed the capabilities of the machine they are running on; but they can all have access, with relatively modest overhead, to the same device.
I don't know if things work as pleasantly as desired yet; but in principle it should be a lot more convenient than full device passthrough. Especially in cases where you might be interested in the GPU for its computational capabilities, video transcoder, etc.