Node restarts when it "wants"

From my observations, ext4 metadata cache for typical Storj node files take around 1 GB of RAM per 1 TB of stored data. A theoretical estimate on this based on the size of metadata (around 300 bytes per file: inode + direntry, around 1 M to 3 M files per 1 TB, plus some overhead for data not laid out in contiguous way).

You can make it more efficient if you reduce the size of an inode at file system creation time—though this parameter cannot be changed if you already have data on your file system. If you don’t have enough RAM, you should look for some other means of caching metadata, for example an SSD-based cache, otherwise you risk running out of IOPS as your node grows.