A web app for sizing an NVMe cache, an HDD tier, and a tape carousel together. It answers whether tape can take archive and recall traffic so the HDD tier can shrink below a 2.5 EB baseline and still serve the compute workload.
go run .
The simulator listens on port 8090 on all interfaces. Open http://127.0.0.1:8090 on this machine. Pass -addr to use another address.
NVMe and HDD. Nodes, a 100/400/800 GbE network, drives per node, drive size, sequential bandwidth, and IOPS. Delivered bandwidth is the minimum of aggregate drive bandwidth and the node network. Network GB/s is the line rate divided by 8.
HDD streams. Active read and write streams share each drive. Their sum divided by the drive count is n in
BW(n) = n·S / (n·(S / BW_seq) + (n − 1)·t_seek)
for more than one stream per drive. Below that, seeks are not charged. The chart holds the read/write mix and sweeps n, with bandwidth on the left axis and the fraction of sequential bandwidth lost to seeks on the right.
Tape. Bandwidth is the drive count times bandwidth per drive. Library capacity is a separate input and does not follow from the drive count.
Workload. Compute read and write, an NVMe hit rate, working set, reserve, archive in EB/year, recall, and an observed TB/s-per-EB reference. The size sweep changes only the HDD node count and reports bandwidth, slack, and cost at each size.
Cost. A node price for every NVMe and HDD server, media in CHF/TB, a price per tape drive, and tape media in CHF/TB. Each tier is split into servers and media.
Units are decimal: 1 EB = 1,000,000 TB. A year is 365 days.
go test ./...