On Thu, Feb 14, 2013 at 02:31:38PM +0100, Paolo Bonzini wrote: > Il 14/02/2013 14:24, Eduardo Habkost ha scritto: > > On Thu, Feb 14, 2013 at 01:13:18PM +0100, Paolo Bonzini wrote: > >> Il 14/02/2013 12:18, Eduardo Habkost ha scritto: > >>>>> qemu boots from disk image 3 times faster than direct kernel load. > >>> That's surprising. Do you have any idea why that happens? > >> > >> Because kernel load uses MMIO (from fw_cfg), while booting from disk > >> uses at worst PCI DMA and at best virtio. > > > > Is it something worth trying to optimize > > I think that, within the limits of what the spec makes legal, Gleb > optimized all that he could out of it. The alternative is to make > fw_cfg do DMA, which in the past was rejected because it doesn't look > like what real ISA hardware would do. > > > , or a reasonable solution would > > be so similar to having a disk+bootloader that's easier to simply > > recommend people to set up a real disk with a real bootloader if they > > care about speed? > > In the end it's a pity, but yeah that's the easiest thing to do with > distro kernels and big all-drivers initrd. -kernel is still useful and > fast enough if you have a custom-built kernel, possibly with no initrd > at all.
The patch that originated this thread wasn't even for distro kernels and big initrds. Our test case lods a very small test kernel (11 KB), and it is taking almost 15 seconds to boot. Maybe our test case should create a custom BIOS image to be loaded on ROM, instead? -- Eduardo
