Dave McMurtrie wrote: > Though I have no experience with it, I seem to recall that someone > attempted to use GFS with an active/active Cyrus cluster and it was a > disaster. It was mentioned either on info-cyrus or in the Cyrus wiki. > If google doesn't help you find this, I can try to remember where I read it. > Replying to myself: Close... It was GPFS, not GFS. Here's the info-cyrus post I was thinking of: http://lists.andrew.cmu.edu/pipermail/info-cyrus/2004-February/008975.html A couple points that Mr. Ulmer makes are valid. When we first went into production with the active/active cluster at Pitt, it ground to a halt under heavy load because of mmap() usage. We ended up changing the map_refresh() code so it would MAP_PRIVATE instead of MAP_SHARED. This relieved the cluster of the burden of tracking the mmap() state across all of the nodes. Also, we did not use Berkeley DB or Skiplist for anything. All of our databases were flat. I think BDB may use shared memory, which definitely won't work across cluster nodes using only a distributed filesystem. If this is the case, BDB just can't be used. HTH, Dave -- Dave McMurtrie, SPE Email Systems Team Leader Carnegie Mellon University, Computing Services ---- Cyrus Home Page: http://cyrusimap.web.cmu.edu/ Cyrus Wiki/FAQ: http://cyrusimap.web.cmu.edu/twiki List Archives/Info: http://asg.web.cmu.edu/cyrus/mailing-list.html
