Crusoe says zstd cuts cold image pulls 2.4 to 3.0 times versus gzip. No venue prices that claim, odds are absent, volume is zero and nothing moved. That absence is the tell for a benchmark with no adoption market. Not financial advice. Markets move fast.
The obvious read is convert to zstd and pocket minutes on every cold start.
I doubt it is that clean. Crusoe said a 3.6 GB CUDA image fell from 51.5 seconds to 17.4 seconds after conversion in its managed Kubernetes tests. Crusoe said a 24.3 GB NeMo image fell from 394 seconds to 152 seconds in the same three-trial cold-pull setup. Crusoe put pod startup gains at 2.2 to 2.4 times in fresh-pod tests with caches cleared.
Crusoe is right about where time goes. Crusoe measured unpack at about 80 percent of pull time on 7 GB and 13 GB images, with containerd unpacking one layer at a time by default. Crusoe said moving the registry in-region saved only 8 percent, 51.5 seconds versus 56.2 seconds from Nvidia public. Compression is the bottleneck.
The conversion tax lands on push and provenance. Crusoe said the Go encoder in skopeo and BuildKit compresses near 180 MB/s while the native zstd CLI hits near 620 MB/s on the same data, so bulk work needs precompressed blobs. Crusoe timed conversion of a 7 GB image near 4 minutes end to end. Crusoe said each convert mints new digests, so pins, signatures and allow-listed digests must be updated.
Runtime support is broad but not free. Crusoe said any node on containerd 1.5.0 or newer unpacks zstd with no changes. Crusoe said Docker before v23 with classic storage fails with invalid tar header, and Slurm with Pyxis/enroot needs enroot 4.0.0 or newer. Crusoe said pull-through caches cannot transcode because recompression breaks digests.
Parallel unpack is the second lever, and it is version gated. Crusoe said raising downloads alone changed nothing because containerd max_concurrent_unpacks defaults to 1 and does not exist before containerd 2.2. Crusoe said 8 downloads with 8 unpacks cut a 7.4 GB zstd cold pull from 82 seconds to 28 seconds. A Docker Compose report with the containerd image store shows the risk in the other direction, with 21 parallel unpacks of a shared 5.7 GB image on Jetson Thor logging about 17,600 Write failed errors, per the issue tracker.
Stacked, the numbers hold. Crusoe said a 13 GB image fell from 183.8 seconds on gzip defaults to 41.5 seconds on tuned zstd, a 4.43 times gain, with format alone at 1.94 times and tuning alone at 2.42 times. Crusoe measured raw decompression at 681 MB/s versus 95 MB/s for gzip, about 7.2 times, with 4 times less CPU and images about 18 percent smaller. Crusoe said below about 100 MB gains fade to fixed overhead.
Convert your three largest images, re-sign them, then touch concurrency.
