Jody McIntyre
2007-Sep-18 23:43 UTC
[Lustre-discuss] WARNING: data corruption for > 2TB OSTs on 32-bit systems
Bug 13627 was recently discovered with a test Lustre installation that resulted in silent data corruption. If a large (more than 2TB) OST is set up on a 32-bit kernel the disk offset for the data IO (though not the filesystem metadata) would be wrapped at 2TB and possibly overwrite existing data in the filesystem. This problem is not present on 64-bit kernels or on OSTs that are smaller than 2TB in size. Note that the word size of the kernel is important, not the CPU it is running on, so a 32-bit (i386, i686) kernel on a 64-bit CPU (x86_64) with larger than 2TB OSTs is affected by this problem. Since this problem is always present, it would normally have shown itself right after installation, and virtually all new Lustre installations (which are likely to have larger OSTs) are using 64-bit kernels for improved performance and large memory management. This problem will be fixed in 1.4.12, and 1.6.3 Lustre releases. Until these are released, you should not create OSTs larger than 2TB on 32-bit kernels. If you believe your existing filesystem is affected by this issue, please contact CFS support (http://bugzilla.lustre.org). There are no known instances of this problem on production systems. -- The Lustre Team --