Re:Show HN: Katamaran - Open-source Docker and Linux runtime...
by nklmilojevic on Friday November 20, 2026 @03:50PM
(#50581863)
Hey HN,
I built Katamaran, native Apple Sillicon Docker and Linux runtime. Most of the Docker tools use super slow Docker Engine and they become annoying to use - f.e. you would feel it on real bind-mount workload like node_modules or PHP's vendor and composer with trees of tens of thousands of small files.
The only product that I know of that fixes this is Orbstack, but it is not opensource, and you can only use it for personal usage.
That is where Katamaran comes - it is even faster than Orbstack, but open-source so you can use it both for personal and business usage.
Why it is faster than others (and comparable or faster than Orbstack - you can check the benchmarks on https://github.com/nklmilojevic/katamaran where it is faster):
- Bind mounts - it based on a patched libkrun over Hypervisor.framework. SMall and boudned FUSE ops run inline on the vCPU via an ARM SMCCC hypercall instead of bouncing through a virtqueue and workter thread. Ofc, it also uses APFS, single FSEvents stream over vsock and in-container inotify
- It also uses patched gvproxy in process inside the VMMM, and resolves DNS via macOS dns_sd
- Reclaiming of memory is super fast becuase it maps guest RAM in 64 kiB XNU VM so virtio-balloon frees return to macOS immediatelly.
In side-by-side benchmarks against OrbStack 2.2.5 on an M3 Max (harness and raw numbers are in bench/), the largest differences are VM stop (0.12s vs 1.95s), uncached DNS across 8 hosts (0.56ms vs 8.57ms; cached 0.09ms vs 0.92ms), host-write to container inotify latency (13.6ms vs 75ms), Laravel db:seed (7.59s vs 27.5s), warm git status on the Linux tree over a bind mount (0.17s vs 0.41s), find over node_modules (0.11s vs 0.22s), and memory footprint 10s after a 4 GiB container exits (2,769 MiB vs 5,632 MiB).
I intentionally built a real use case benchmark rather than some synthetic micro-benchmarks that wouldn't show actual performance.
I'm open for feedback, please let me know how it goes for you :)
I built Katamaran, native Apple Sillicon Docker and Linux runtime. Most of the Docker tools use super slow Docker Engine and they become annoying to use - f.e. you would feel it on real bind-mount workload like node_modules or PHP's vendor and composer with trees of tens of thousands of small files.
The only product that I know of that fixes this is Orbstack, but it is not opensource, and you can only use it for personal usage.
That is where Katamaran comes - it is even faster than Orbstack, but open-source so you can use it both for personal and business usage.
Why it is faster than others (and comparable or faster than Orbstack - you can check the benchmarks on https://github.com/nklmilojevic/katamaran where it is faster):
- Bind mounts - it based on a patched libkrun over Hypervisor.framework. SMall and boudned FUSE ops run inline on the vCPU via an ARM SMCCC hypercall instead of bouncing through a virtqueue and workter thread. Ofc, it also uses APFS, single FSEvents stream over vsock and in-container inotify - It also uses patched gvproxy in process inside the VMMM, and resolves DNS via macOS dns_sd - Reclaiming of memory is super fast becuase it maps guest RAM in 64 kiB XNU VM so virtio-balloon frees return to macOS immediatelly.
In side-by-side benchmarks against OrbStack 2.2.5 on an M3 Max (harness and raw numbers are in bench/), the largest differences are VM stop (0.12s vs 1.95s), uncached DNS across 8 hosts (0.56ms vs 8.57ms; cached 0.09ms vs 0.92ms), host-write to container inotify latency (13.6ms vs 75ms), Laravel db:seed (7.59s vs 27.5s), warm git status on the Linux tree over a bind mount (0.17s vs 0.41s), find over node_modules (0.11s vs 0.22s), and memory footprint 10s after a 4 GiB container exits (2,769 MiB vs 5,632 MiB).
I intentionally built a real use case benchmark rather than some synthetic micro-benchmarks that wouldn't show actual performance.
I'm open for feedback, please let me know how it goes for you :)