The comparison with QEMU is a bit disingenuous. The number of lines of code in a given binary is much less than 1.4 million, which is the total for all the architectures that QEMU supports (actually it's closer to 2 million). It is also possible to configure out a lot of the code and libraries. A default build is around 6-700.000 lines of code, and it is possible to build reduced versions that tally less than 300.000 lines just by tweaking a configuration file with the list of included devices (which was done for the experiments in the paper). The NEMU project mentioned in the paper is inactive exactly because it was a dead end: it was possible to achieve all their goals directly in QEMU without the need for a fork.<p>Likewise, the number of syscalls in QEMU (270) is quoted from another paper but likely refers not to QEMU used as a VMM, but rather to the so called "user-mode emulation" that runs foreign Linux binaries (and which, by its very nature, invokes pretty much all Linux syscalls including quite a few obsolete ones).<p>It would be interesting to have more information on the configuration they use for QEMU, especially whether they are enabling PCI and ACPI. Version 4.2 of QEMU (the version they used) has a trimmed virtual machine type that was admittedly inspired by Firecracker and actually boots even faster than Firecracker, or at least in the noise. [1] The code specific to this machine type is only 500 lines. Also since that release QEMU and qboot _are_ actually able to boot uncompressed kernels using the PVH entry point, contrarily to what the paper states. This could explain the difference in boot time performance.<p>Memory consumption is tricky, QEMU uses green threads and has to allocate stacks for them. This can show up as large mmaps but they do not correspond to actual memory usage. But again without redoing the experiments it's hard to say if that is the cause. I have no doubt Firecracker is leaner in this respect, anyway, and it's an important metric for Amazon.<p>That said, I am all for competition, and some of the measurements in the paper are certainly worth a look to see if there is more low hanging fruit to pick, especially with respect to memory consumption. QEMU is a complex program, as demonstrated by the complexity of measuring it accurately, and Firecracker is a very interesting project. I am very happy to collaborate with the authors of the paper on rust-vmm.<p>[1] <a href="http://people.redhat.com/pbonzini/microvm/" rel="nofollow">http://people.redhat.com/pbonzini/microvm/</a>