hw/i386/pc.c | 15 ++++++++--- include/hw/elf_ops.h | 64 ++++++++++++++++++++++++-------------------- 2 files changed, 46 insertions(+), 33 deletions(-)
In order to reduce the memory footprint when PVH kernel and initrd
are used, we map them into memory instead of reading them.
In this way we can share them between multiple instances of QEMU.
v2:
- Patch 1: used g_mapped_file_new_from_fd() with 'writeble' set to 'true',
since we can modify the mapped buffer. [Paolo, Peter]
These are the results using a PVH kernel and initrd (cpio):
- memory footprint (using smem) [MB]
QEMU before now
# instances USS PSS USS PSS
1 102.0M 105.8M 102.3M 106.2M
2 94.6M 101.2M 72.3M 90.1M
4 94.1M 98.0M 72.0M 81.5M
8 94.0M 96.2M 71.8M 76.9M
16 93.9M 95.1M 71.6M 74.3M
Initrd size: 3.0M
Kernel
image size: 28M
sections size [size -A -d vmlinux]: 18.9M
- boot time [ms]
before now
qemu_init_end: 63.85 55.91
linux_start_kernel: 82.11 (+18.26) 74.51 (+18.60)
linux_start_user: 169.94 (+87.83) 159.06 (+84.56)
QEMU command used:
./qemu-system-x86_64 -bios /path/to/seabios/out/bios.bin -no-hpet \
-machine q35,accel=kvm,kernel_irqchip,nvdimm,sata=off,smbus=off,vmport=off \
-cpu host -m 1G -smp 1 -vga none -display none -no-user-config -nodefaults \
-kernel /path/to/vmlinux -initrd /path/to/rootfs.cpio \
-append 'root=/dev/mem0 ro console=hvc0 pci=lastbus=0 nosmap'
Stefano Garzarella (2):
elf-ops.h: Map into memory the ELF to load
hw/i386/pc: Map into memory the initrd
hw/i386/pc.c | 15 ++++++++---
include/hw/elf_ops.h | 64 ++++++++++++++++++++++++--------------------
2 files changed, 46 insertions(+), 33 deletions(-)
--
2.20.1
Stefano, Brilliant job! I can confirm that with these patches the memory footprint is smaller and the boot time is the same for kata Here the results using kata metrics https://pasteboard.co/Ipl06Q0.png https://pasteboard.co/Ipl3p4d.png Thanks - Julio On Tue, 2019-07-23 at 16:04 +0200, Stefano Garzarella wrote: > In order to reduce the memory footprint when PVH kernel and initrd > are used, we map them into memory instead of reading them. > In this way we can share them between multiple instances of QEMU. > > v2: > - Patch 1: used g_mapped_file_new_from_fd() with 'writeble' set to > 'true', > since we can modify the mapped buffer. [Paolo, Peter] > > These are the results using a PVH kernel and initrd (cpio): > - memory footprint (using smem) [MB] > QEMU before now > # instances USS PSS USS PSS > 1 102.0M 105.8M 102.3M 106.2M > 2 94.6M 101.2M 72.3M 90.1M > 4 94.1M 98.0M 72.0M 81.5M > 8 94.0M 96.2M 71.8M 76.9M > 16 93.9M 95.1M 71.6M 74.3M > > Initrd size: 3.0M > Kernel > image size: 28M > sections size [size -A -d vmlinux]: 18.9M > > - boot time [ms] > before now > qemu_init_end: 63.85 55.91 > linux_start_kernel: 82.11 (+18.26) 74.51 (+18.60) > linux_start_user: 169.94 (+87.83) 159.06 (+84.56) > > QEMU command used: > ./qemu-system-x86_64 -bios /path/to/seabios/out/bios.bin -no-hpet \ > -machine > q35,accel=kvm,kernel_irqchip,nvdimm,sata=off,smbus=off,vmport=off \ > -cpu host -m 1G -smp 1 -vga none -display none -no-user-config > -nodefaults \ > -kernel /path/to/vmlinux -initrd /path/to/rootfs.cpio \ > -append 'root=/dev/mem0 ro console=hvc0 pci=lastbus=0 nosmap' > > Stefano Garzarella (2): > elf-ops.h: Map into memory the ELF to load > hw/i386/pc: Map into memory the initrd > > hw/i386/pc.c | 15 ++++++++--- > include/hw/elf_ops.h | 64 ++++++++++++++++++++++++---------------- > ---- > 2 files changed, 46 insertions(+), 33 deletions(-) >
On Tue, Jul 23, 2019 at 05:37:18PM +0000, Montes, Julio wrote: > Stefano, Brilliant job! > > I can confirm that with these patches the memory footprint is smaller > and the boot time is the same for kata > > Here the results using kata metrics > > https://pasteboard.co/Ipl06Q0.png > https://pasteboard.co/Ipl3p4d.png > Hi Julio, thank you very much for testing :) When you measure the PPS, how many QEMU instances did you start? Did you use also the initrd? Thanks, Stefano
Hi Stefano On Wed, 2019-07-24 at 09:36 +0200, Stefano Garzarella wrote: > On Tue, Jul 23, 2019 at 05:37:18PM +0000, Montes, Julio wrote: > > Stefano, Brilliant job! > > > > I can confirm that with these patches the memory footprint is > > smaller > > and the boot time is the same for kata > > > > Here the results using kata metrics > > > > https://pasteboard.co/Ipl06Q0.png > > https://pasteboard.co/Ipl3p4d.png > > > > Hi Julio, > thank you very much for testing :) > > When you measure the PPS, how many QEMU instances did you start? > Did you use also the initrd? 50 VMs with a nvdimm/pmem device as rootfs, I will test your v3 with initrd :) > > Thanks, > Stefano
On Wed, Jul 24, 2019 at 01:03:42PM +0000, Montes, Julio wrote: > Hi Stefano > > On Wed, 2019-07-24 at 09:36 +0200, Stefano Garzarella wrote: > > On Tue, Jul 23, 2019 at 05:37:18PM +0000, Montes, Julio wrote: > > > Stefano, Brilliant job! > > > > > > I can confirm that with these patches the memory footprint is > > > smaller > > > and the boot time is the same for kata > > > > > > Here the results using kata metrics > > > > > > https://pasteboard.co/Ipl06Q0.png > > > https://pasteboard.co/Ipl3p4d.png > > > > > > > Hi Julio, > > thank you very much for testing :) > > > > When you measure the PPS, how many QEMU instances did you start? > > Did you use also the initrd? > > 50 VMs with a nvdimm/pmem device as rootfs, I will test your v3 with > initrd :) Cool :) Okay, I'm sending a v4 to fix some issues... Thanks, Stefano
© 2016 - 2024 Red Hat, Inc.