Re: Revisiting parallel save/restore

On Fri, Apr 26, 2024 at 11:44:38AM -0300, Fabiano Rosas wrote: > Daniel P. Berrangé <berrange(a)redhat.com&gt; writes: > >> On Fri, Apr 26, 2024 at 10:03:29AM -0300, Fabiano Rosas wrote: >>> Daniel P. Berrangé <berrange(a)redhat.com&gt; writes: >>> >>>> On Wed, Apr 17, 2024 at 05:12:27PM -0600, Jim Fehlig via Devel wrote: >>>>> A good starting point on this journey is supporting the new mapped-ram >>>>> capability in qemu 9.0 [2]. Since mapped-ram is a new on-disk format, I >>>>> assume we'll need a new QEMU_SAVE_VERSION 3 when using it? Otherwise I'm not >>>>> sure how to detect if a saved image is in mapped-ram format vs the existing, >>>>> sequential stream format. >>>> >>>> Yes, we'll need to be supporting 'mapped-ram', so a good first step. >>>> >>>> A question is whether we make that feature mandatory for all save images, >>>> or implied by another feature (parallel save), or an directly controllable >>>> feature with opt-in. >>>> >>>> The former breaks back compat with existnig libvirt, while the latter 2 >>>> options are net new so don't have compat implications. >>>> >>>> In terms of actual data blocks written on disk mapped-ram should be be the >>>> same size, or smaller, than the existing format. >>>> >>>> In terms of logical file size, however, mapped-ram will almost always be >>>> larger. >>>> >>>> This is because mapped-ram will result in a file whose logical size matches >>>> the guest RAM size, plus some header overhead, while being sparse so not >>>> all blocks are written. >>>> >>>> If tools handling save images aren't sparse-aware this could come across >>>> as a surprise and even be considered a regression. >>>> >>>> Mapped ram is needed for parallel saves since it lets each thread write >>>> to a specific region of the file. >>>> >>>> Mapped ram is good for non-parallel saves too though, because the mapping >>>> of RAM into the file is aligned suitably to allow for O_DIRECT to be used. >>>> Currently libvirt has to tunnnel over its iohelper to futz alignment >>>> needed for O_DIRECT. This makes it desirable to use in general, but back >>>> compat hurts... >>> >>> Note that QEMU doesn't support O_DIRECT without multifd. >>> >>> From mapped-ram patch series v4: >>> >>> - Dropped support for direct-io with fixed-ram _without_ multifd. This >>> is something I said I would do for this version, but I had to drop >>> it because performance is really bad. I think the single-threaded >>> precopy code cannot cope with the extra latency/synchronicity of >>> O_DIRECT. >> >> Note the reason for using O_DIRECT is *not* to make saving / restoring >> the guest VM faster. Rather it is to ensure that saving/restoring a VM >> does not trash the host I/O / buffer cache, which will negatively impact >> performance of all the *other* concurrently running VMs.

> > Well, there's surely a performance degradation threshold that negates > the benefits of perserving the caches. But maybe it's not as low as I > initially thought then. I guess you could say that O_DIRECT makes saving/restoring have a predictable speed, because it will no longer randomly vary depending on how much free RAM happens to be available at a given time. Time will be dominated largely by the underlying storage I/O performanc

With regards, Daniel