Skip to content

Benchmarks

These measurements come from run 20261120c on a Mac15,10 with an Apple M3 Max (14 cores, 10 performance and 4 efficiency, 36 GiB RAM, macOS 27.2.1). The katamaran VM used 14 vCPUs, a 16 GiB memory ceiling, kernel 6.18.64, Docker 29.10.0, and an Alpine 3.24 rootfs. Every number is the median of 5 timed runs unless noted.

Workload Median
stat() on a 91,000-file Linux tree (warm) 346 ns per file
stat() on a 91,000-file Linux tree (guest caches dropped) 6.44 µs per file
open() and close() (warm) 24.5 µs per pair
find . -type f over node_modules (33,598 files) 0.113 s
git status on the Linux tree (warm) 0.166 s
git status on the Linux tree (guest caches dropped) 6.43 s
git status in a new container right after cold pass 0.176 s
tar -xf of linux-6.18.tar.xz onto a bind mount 13.8 s
rm -rf of the extracted Linux tree on a bind mount 5.71 s
npm ci on video.js (warm cache) 3.23 s
pnpm install on video.js (warm store) 2.32 s
cargo build on ripgrep (debug, offline) 3.51 s
Host file write to container inotify event 13.6 ms
Inotify events lost in a 10,000-file burst 0
fio sequential read, 1 MiB blocks 4,092 MB/s
fio sequential write, 1 MiB blocks 2,006 MB/s
fio random read, 4 KiB blocks 1,098 MB/s (267,965 IOPS)
fio random write, 4 KiB blocks 747 MB/s (182,430 IOPS)
fio fsync write, 4 KiB blocks 26.9 MB/s (6,572 IOPS)

Measured with a Laravel 12 and Vite application on a bind mount plus PostgreSQL 17 and Redis 7 containers:

Workload Median
composer install (warm cache) 3.54 s
bun install (warm cache) 3.95 s
vite build 6.26 s
php artisan migrate:fresh (Postgres 17) 0.772 s
php artisan db:seed total (12,000 inserts and 3,000 row-lock transactions) 7.59 s
Laravel GET /login, opcache off 155 ms per request
Laravel GET /login, opcache on 26.3 ms per request
10 parallel unzips of the Composer cache 0.981 s
Container to container Redis PING round trip (redis-benchmark -c 1 -n 20000) 6.15 µs
Workload Median
Container downloads 1 GiB from macOS (host.docker.internal) 4,440 MiB/s
Container uploads 1 GiB to macOS 5,698 MiB/s
macOS downloads 1 GiB from a published port (-p) 5,131 MiB/s
HTTP to macOS, keep-alive p50 and p99 95.6 µs (p99 151 µs)
HTTP to macOS, new connection p50 and p99 334 µs (p99 651 µs)
DNS getaddrinfo for one cached name 91.5 µs
DNS getaddrinfo for 8 different hosts 0.56 ms

Linux to macOS files (~/Katamaran over NFS)

Section titled “Linux to macOS files (~/Katamaran over NFS)”
Workload Median
Read a 100 MiB file on macOS 981 MiB/s
Write a 100 MiB file from macOS 1,090 MiB/s
Read 10,000 files of 4 KiB (cat) 3.55 s
ls -R over a 21,000-file tree 4.08 s
cp -R of a 21,000-file tree to macOS 12.8 s
Workload (linux/amd64) Median
200 executions of /bin/true 495 ms
Python sum(i*i) over 10,000,000 integers 849 ms
Node loop of 300,000,000 iterations 1,179 ms
gcc -O2 compiling 3,000 functions 20.2 s
10 executions of node -e 0 3.30 s

Physical footprint is measured from vmmap -summary on the VM processes.

Measurement Median
Cold start to docker ready 668 ms
Cold start to first docker run --rm alpine true 967 ms
Stop VM (SIGINT to process exit) 120 ms
docker run --rm alpine true (10 runs) 302 ms
Idle footprint on fresh VM, no containers 922 MiB
Idle footprint after dropping guest caches (fresh VM) 895 MiB
Idle CPU, no containers (5 windows of 30 s) 1.13 % to 1.47 % of one core
Peak footprint with a 4 GiB container running 5,636 MiB
Footprint 10 s after the 4 GiB container exits 2,769 MiB
Footprint 60 s after the 4 GiB container exits 1,540 MiB

The benchmark tool in bench/ uses uv and Python 3. Prepare the datasets once, then run all or selected cases against your katamaran context:

Terminal window
cd bench
uv run bench prepare --engine katamaran
uv run bench run --engine katamaran --locations bind --runs 5

List available cases:

Terminal window
uv run bench list

Measure network throughput and HTTP latency directly:

Terminal window
bash scripts/netperf.sh katamaran