4 ms·
I like having separate partitions (or slices) for everything. The guy who introduced me to UNIX did it so he could mount certain filesystems "ro" or "noexec". H
by Esau 10y ago
I like having separate partitions (or slices) for everything. The guy who introduced me to UNIX did it so he could mount certain filesystems "ro" or "noexec". He also told me that partitioning can help avoid inode exhaustion but I really doubt that is an issue with modern filesystems.
I still partition with NetBSD. It just feels right; even if not necessary.
- protomyth 10y agoOh, I partition too, but I do /, /usr, /var, /var/log, /home, /tmp, /opt (yeah, yeah, I know), and then anything server specific (mail, maybe www). I just don't do the /usr/local.
- Esau 10y agoWhen I set up NetBSD, I create /, /usr, /var. /tmp, and /home. I have never create a /usr/local partition but I can see value in it is you want /usr itself to be read-only.
- protomyth 10y agoI do /var/log because of some bad experiences with runaway programs - samba went psycho on me in the 3.x days.
- Tharkun 10y agoI never run out of disk space, but I frequently run out of inodes. It's definitely still an issue. And if you don't stop to think about when you're creating your file system, it tends to bite you in the arse at a later date.
- Esau 10y agoOut of curiosity, can you share what platform/file system you generally use? I had assumed inodes would not be an issue with modern operating systems; especially 64-bit ones.
- kazinator 10y agoOn traditional Unix filesystems, a fixed number of inodes is provisioned when a filesystem is created. Though of course the upper bound on that is limited by the width of the numeric type ino_t (or whatever plays that role in kernel space) have you, in practice, the actual limit is lower. I'm looking at a filesystem I have handy here; the root FS in an Ubuntu Linux VM. dumpe2fs /dev/sda1 says that the inode count is 1146880. Just a million and something: way less than what a even a 32 bit inode number can represent. That filesystem simply cannot have more objects than that, in total. With this type of filesystem, you have to estimate how many objects you will ever need to store in it, not only how much total storage. If you estimate too high, you waste inodes. Suppose inodes are 128 bytes wide and you provision four billion of them. Oops, that's like half a terabyte of storage dedicated to the array of inodes!