(thanks for the ref to the neu.edu paper, I hadn't read that.)

On Sep 24, 2013, at 6:07 PM, Stefan Monnier wrote:

> - I've found USB flash disks to be reasonably reliable in general, but
>  one case where I've had trouble is when disconnecting the key while
>  the system is writing to it.  That can lead to severe corruption
>  (despite use of things like logging file-systems).

I've seen that, except without the disconnect. 

USB Flash disks have no consistency/atomicity guarantees, at all. If there is a 
failure of any sort, it may result in "impossible" filesystem contents, and 
fsck.ext3 may be extremely confused. Many consumer flash drives are optimized 
for a FAT and may have different commit behavior for blocks below and above the 
(say) 4M mark.

I stopped trying to run 24/7 systems with / on USB Flash about five years ago. 
Things *may* have improved since then, but I wouldn't know....

If you can live with a livecd-like cramfs overlay, that might work pretty well. 
Just be prepared to lose the overlay filesystem, or anything else you touch 
frequently. If somebody held a gun to my head and said "build me a reliable 
system" I would probably keep tgz snapshots of /overlay on a FAT filesystem; on 
boot, I would restore the last one with a good checksum. Some people say "don't 
ever reformat" because of alignment issues etc, but it's probably safe to 
shrink the main FAT filesystem and keep the cramfs on a separate partition at 
the end of the drive.

If USB ports are not a problem, using a second drive might help a lot. The 
standard performance disaster is demand-loading an executable from /usr/bin 
while the (queue-depth 0) disk is frozen with 10IOPS writes to /var/log or 
$HOME/.browsercache/.

Jay
_______________________________________________
openwrt-users mailing list
[email protected]
https://lists.openwrt.org/cgi-bin/mailman/listinfo/openwrt-users

Reply via email to