TensorDock
Search…
⌃K

Storage

Our storage classes

General Info

When you deploy a TensorDock server, by default, the storage will be a network-based volume. Our storage arrays are:
  • 3x-replicated, so your data won't be deleted
  • Ceph-based, so it's reliable
  • Backed by 10/40/100 Gbit internal networking (depending on location)

Cheap, Fast, Durable

We value your data. We are honored to have you trust us with your data. At the same time, data loss is possible, even at big providers, so we recommend that you follow the 3-2-1 backup rule.
  • Cheap: our storage volumes cost just more than $0.07/GB/month, which is cheaper than AWS and even DigitalOcean (as of January 28th, 2022)
  • Fast: you should be able to get above 1 GB/s on our NVMe-based network SSDs with 1 MB block sizes
  • Durable: with our triple-replicated setup, you can feel confident that hard drive failures won't take your data down

Storage Classes

Performance-Optimized NVMe SSD Storage (io1)

Cost: $0.0001/GB/hour ($0.073/GB/month, or $73.00/TB/month)
Our io1 class of NVMe-based network storage is our flagship storage class. These are NVMe-based SSDs, networked together for maximum random read and write speeds.
Here is the disk performance excerpt from a YABS benchmark of a Quadro RTX 4000, 2 vCPU, 4GB RAM, 20 GB io1 class storage machine:
fio Disk Speed Tests (Mixed R/W 50/50):
---------------------------------
Block Size | 4k (IOPS) | 64k (IOPS)
------ | --- ---- | ---- ----
Read | 92.50 MB/s (23.1k) | 965.67 MB/s (15.0k)
Write | 92.74 MB/s (23.1k) | 970.75 MB/s (15.1k)
Total | 185.24 MB/s (46.3k) | 1.93 GB/s (30.2k)
| |
Block Size | 512k (IOPS) | 1m (IOPS)
------ | --- ---- | ---- ----
Read | 1.03 GB/s (2.0k) | 1.03 GB/s (1.0k)
Write | 1.08 GB/s (2.1k) | 1.10 GB/s (1.0k)
Total | 2.12 GB/s (4.1k) | 2.13 GB/s (2.0k)

Performance-Optimized Hard Drive Storage (st1)

Cost: $0.00006/GB/hour ($0.0438/GB/month, or $43.80/TB/month)
Our st1 class of HDD-based network storage is perfect for performant data storage. These are hard drives, networked together for maximum sequential read and write speeds.
Here is the disk performance excerpt from a YABS benchmark of a Quadro RTX 4000, 2 vCPU, 4GB RAM, 20 GB st1 class storage machine:
fio Disk Speed Tests (Mixed R/W 50/50):
---------------------------------
Block Size | 4k (IOPS) | 64k (IOPS)
------ | --- ---- | ---- ----
Read | 44.69 MB/s (11.1k) | 328.41 MB/s (5.1k)
Write | 44.76 MB/s (11.1k) | 330.13 MB/s (5.1k)
Total | 89.46 MB/s (22.3k) | 658.54 MB/s (10.2k)
| |
Block Size | 512k (IOPS) | 1m (IOPS)
------ | --- ---- | ---- ----
Read | 903.52 MB/s (1.7k) | 951.06 MB/s (928)
Write | 951.53 MB/s (1.8k) | 1.01 GB/s (990)
Total | 1.85 GB/s (3.6k) | 1.96 GB/s (1.9k)

I/O-Optimized vs Storage-Optimized

Our io1 storage class is optimized for random reads and writes and is the faster storage medium, whereas our st1 storage class is optimized for more capacity at the same cost.