...
- UNC's cluster is up and running, some details added to the wiki
- Was using file stores for back end storage, ran several performance tests (though not with a larger number of files)
- behind load balancer
- when not using transactions, everything was fine
- transactions didn't work...
- possible solutions include:
- hacking sessions using cookies or fancy routing and rewrite rules
- use JTA transactions rather than sharing a session
- hitting the first node in the cluster
- transactions and threads all worked
- behind load balancer
- eventually things broke and everything returns an error
- Is the error reproducablereproducible? (Andrew)
- Yes... in that the errors are constant now (possibly when using transaction)
- Was using file stores for back end storage, ran several performance tests (though not with a larger number of files)
Andrew summarized... number of issues around clustering
- transactions
- fixity (when segementedsegmented)
- setting it up is still hard (Esme, Scott are having trouble just getting it up)
- Greg identified TCP issues and documented them, likely making future setups easier
- Gret will reconfigure his cluster for levelDB, then go from TCP to UDP (which is more efficient)
- Greg asked whether the levelDB storage is supposed to be shared or independent between nodes, assuming the later.
...
Andrew wonders about the theory that performance could improve without sending requests to different nodes.
Frank found performance was worse in a previous test using a load-balancer
Puppet-fcrepo planning
Andrew wants us to get to a place where we have scripts (puppet) to easily bring up a cluster as a prequisite prerequisite for more cluster testing.
...