Opened 10 years ago
Collective replication over GM requires "server-less" connections
|Reported by:||hauma||Owned by:||hauma|
In particular collective updates (see: ticket:157) are much slower when using KaRMI/GM instead of KaRMI/Ethernet. The reason seems to be that collective operations involve lots of (server-) threads for accumulating the required connections. Currently, a collective operation is mapped to multiple connections that normally serve as transport for remote method invocations. Each such connection is associated on one end with a server thread that is either responsible for executing the remote method invocation, or for forwarding the call to a waiting segment of a transparent distributed thread (see ticket:140).
The problem could be mitigated by introducing a second connection abstraction in KaRMI that is not automatically associated with a server thread. Such server-less connection can only be used, if two threads on both ends agree on a collective operation.