Re: [libvirt] [PATCH 01/12] conf: introduce domain XML element <polling> for iothread

On Tue, Feb 21, 2017 at 12:26:02PM +0000, Daniel P. Berrange wrote: > On Tue, Feb 21, 2017 at 01:14:57PM +0100, Pavel Hrdina wrote: > > QEMU 2.9.0 will introduce polling feature for AioContext that instead > > of blocking syscalls polls for events without blocking. This means > > that polling can be in most cases faster but it also increases CPU > > utilization. > > > > To address this issue QEMU implements self-tuning algorithm that > > modifies the current polling time to adapt to different workloads > > and it can also fallback to blocking syscalls. > > > > For each IOThread this all is controlled by three parameters, > > poll-max-ns, poll-grow and poll-shrink. If parameter poll-max-ns > > is set to 0 it disables the polling, if it is omitted the default > > behavior is used and any value more than 0 enables polling. > > Parameters poll-grow and poll-shrink configure how the self-tuning > > algorithm will adapt the current polling time. If they are omitted > > or set to 0 default values will be used. > > With my app developer hat on I have to wonder how an app is supposed > to figure out what to set these parameters to ? It has been difficult > enough figuring out existing QEMU block tunables, but at least most > of those can be set dependant on the type of storage use on the host > side. Tunables whose use depends on the guest workload are harder to > use since it largely involves predicting the unknown. IOW is there > a compelling reason to add these low level parameters that are tightly > coupled to the specific algorithm that QEMU happens to use today. I agree that it's probably way to complicated for management applications, but there is a small issue with QEMU. Currently it you don't specify anything the polling is enabled with some reasonable default value and base on experience with QEMU I'm not planning to count on that they will not change the default behavior in the future. To explicitly enable polling you need to set poll-max-ns to some value more than 0. We would have to check qemu source code, and define the default value in our code in order to let users explicitly enable polling.

> The QEMU commits say the tunables all default to sane parameters so > I'm inclined to say we ignore them at the libvirt level entirely. Yes, it would be way batter to have only <polling enable='yes|no'> end let QEMU deal with the sane values for all parameters but that would mean to come up with the sane values ourselves or modify QEMU to add another property that would simply control whether it is enabled or not.