Re: [SSI] Supporting SSI on environment not having GFS or layered CFS
Brought to you by:
brucewalker,
rogertsang
From: Bruce W. <br...@ka...> - 2001-10-16 18:37:56
|
It certainly hasn't been attempted but I can't think why it wouldn't be possibly with some work. Some of work that might be needed would be: a. create a single node with SSI but with ext2 or some other physical filesystem for the root. b. export the root c. create the modified SSI kernel that can use a remote root over NFS and boot that on another node. d. For other filesystems, make sure they are mounted on the same mount point on all systems e. the "reopen" code that is part of migrate should just work with NFS-client-open files but hasn't been attempted to my knowledge. f. We might want to enhance the process migration to fsync open files before migrating, due to the much looser coherency guarantees of NFS. g. Process migration while holding file locks doesn't currently work. It will be fixed with GFS and CFS. A different fix may be needed over NFS. h. Interesting to see what would happen with fifos. They may just work clusterwide-shared as they do in the current SSI. Some of the limitations of what the result would be: 1. Coherency would be limited to that of NFS. Given process migration, the likelihood of sharing is higher so this is more of a restriction. 2. Basically a single root node, so availability is very poor. Doing NFS to ext2 root filesystem takeover would be quite an exercise. my two cents, bruce > In cluster that keeps only NFS, is the support of SSI > possible? (process migration, clusterwide device space, etc) > > -jlyu |