> Narrowing the interface of a platform will not necessarily improve security, unless it reduces the footprint of host code likely to contain vulnerabilities, so we have focused on understanding the kernel footprint underneath each virtualization system.

These things are hard to quantify, but this is a very reasonable attempt. However, as we say in one of our Firecracker papers (https://www.usenix.org/conference/nsdi20/presentation/agache):

> Li et al [32] demonstrate the effectiveness of a ’popular paths’ metric, showing that only 3% of kernel bugs are found in frequently-used code paths (which, in our experience, overlap highly with the code paths used by the VMM)

There, the reference is to Lock-In-Pop by Li et al (https://www.usenix.org/conference/atc17/technical-sessions/p...). This is important, because bug density isn't constant across the kernel surface area (just as it isn't constant across any code base). FWIW I don't think Lock-In-Pop is the right mechanism for isolation, but it's a very good way to weight the metrics if we're comparing kernel surface area.

Back to the paper at hand:

> Both gVisor and Firecracker restrict the set of system calls available to improve security and isolation.

I'm not sure what they mean here. Firecracker doesn't restrict the set of system calls guests can use. In my mind, that's one of the biggest advantages of Firecracker's approach - there's no hard tradeoff to be made between isolation strength and guest software compatibility.

> I'm not sure what they mean here. Firecracker doesn't restrict the set of system calls guests can use. In my mind, that's one of the biggest advantages of Firecracker's approach - there's no hard tradeoff to be made between isolation strength and guest software compatibility.

gVisor is also not artificially restricting this set (it is limited by what's implemented, but that's not a hard trade-off); the sentence means that both Firecracker and gVisor restrict the host surface available to the effective "supervisor", either the VMM or the Sentry.

+1. I'd also add that gVisor's slightly limited compatibility is a security feature. Like there is no need for obscure functionality which most production apps rarely use.

Note that has been a number of recent performance improvements in gVisor (VFS2, soon LisaFS).

Here's a recent blog post about performance in a large production setup: https://gvisor.dev/blog/2021/12/02/running-gvisor-in-product...

Are firecracker and gvisor available as containers in Ubuntu / Debian?

Follow-up question, can they be implemented in Proxmox? How feasible is it.

I don't know if number of kernel lines executed is a good baseline for establishing the exposure of a given sandboxing approach.

Also, the number of lines of kernel code being executed is not a valid metric without comparing how much of those code flow can the user control. In gVisor, this is mostly limited to what the go runtime can call. There is very little direct control over these syscalls.

Tyler is an absolutely fantastic instructor and mentor! Glad to see his work show up on the front page!