Page fault about firecracker

[thunderZH963]

As firecracker said, it is implemented based on KVM. Then, the processing of page fault should be consistent with KVM / QEMU. But in my test, the delay of page fault in microVM is lower than QEMU / KVM VM. Why?

Hope to get a reply and discussion.

Thanks.

[thunderZH963]

I can observe "let memory_region = kvm_userspace_memory_region{...}" in firecracker/src/vm.rs

or maybe other codes indicate "page fault handler"?

[JonathanWoollett-Light]

But in my test, the delay of page fault in microVM is lower than QEMU / KVM VM. Why?

If you can provide the failing test case we would be interested to take a look.

[thunderZH963]

But in my test, the delay of page fault in microVM is lower than QEMU / KVM VM. Why?

If you can provide the failing test case we would be interested to take a look.

The cases shown in below and the results also are shown.

lat/us	micro fork+exit	micro fork+execve	micro fork+shell	micro pipe	micro mmap-write	micro mmap-read	lmbench fork+exit	lmbench fork+execve	lmbench fork+shell	lmbench pipe	lmbench mmap-write	lmbench mmap-read	lmbench pagefault
Baremetal	93.335	315.87	812.01	1.52	43.177	16.812	98.1964	328.875	807.7143	5.0492	15.823	6.42	0.25
KVM	108.284	540.135	1825.631	1.629	47.932	32.9	113.08	465.1818	1770.3333	5.66	17.022	10.1	1.18
Firecracker	99.57	517.164	1033.429	1.576	43.207	17.36	100.5926	355.625	842.3846	5.1561	17.021	9.56	0.31

[roypat]

Hi @thunderZH963, We do not maintain comparisons with QEMU/other virtualization technologies, so we do not know why you observe these performance differences. Firecracker is more lightweight than QEMU, so I would expect this difference in complexity to play a role. I'm converting this issue to a discussion in case other community members would like to chime in on this :)