Rusty recently committed my patch to convert lguest over to the modern hrtimers API. This allows guests to run tickless (also known as dynamic ticks), which means that instead of being fed a constant stream of timer ticks, a guest will only receive a timer interrupt when it really needs one. As a bonus, these ticks will likely arrive with a much higher resolution than every 1/1000 s (or whatever HZ was set to).

Tickless operation is important for virtualized systems. Imagine if you had ten guests, each running at 1000HZ. With static ticks, the host would need to generate 10,000 virtual ticks per second, and each guest would be handling 1000 synthetic hardware interrupts/s, the vast majority of which would very likely be unnecessary. Instead, lguest now allows guests to request clock ticks to occur only at some appropriate point in the future, with nanosecond resolution.

A virtual clock event device is instantiated for each running guest. Instead of being a programmable piece of hardware (such as an HPET), the virtual device is quite simply a high-resolution timer attached to the guest. When the guest needs to program a clock tick event, the smarts of which are thankfully buried in the hrtimer code, it issues a new hypercall, HCALL_SET_CLOCKEVENT, with an argument in nanoseconds representing the relative amount of time until the event is required. The hypercall traps into the host, which converts the relative delta into an absolute value (we know when ‘now’ is in the guest’s virtual time, as we’re in its hypercall context, which is similar to being in the syscall context of a userland application), then starts a highres timer running in the host. When this fires, the host simply sets the timer interrupt for the guest. When the guest runs next, its own hrtimer mechanism is then invoked to process the timer interrupt.

Conceptually, it’s quite simple, which keeps the puppies happy. This is due in large part to the great hrtimer subsystem by Thomas Gleixner and Ingo Molnar. Check out the 2006 OLS paper for more info on hrtimers, as well as the LWN articles here and here.