Open lukemarsden opened 5 years ago
Update: same test took 13 minutes on prod, perhaps because the prod VM has a better spec. Still, I suspect there's room for improvement by adding mbuffer. See https://everycity.co.uk/alasdair/2010/07/using-mbuffer-to-speed-up-slow-zfs-send-zfs-receive/
in testing, a dataset which took 4 minutes to wget from GCS to a GCP instance at ~200MB/sec, then took 40 minutes to push from GCP runner to GCP hub (staging).
to be fair there was some unzipping involved, which probably 2x'd the dataset size, but it should still be possible to push between dotmesh at disk/network speed.
try adding mbuffer between zfs send/recv to smooth out the stream and increase the throughput.