P5
From a customer a Progress Software (via Matt Hill):
>The sun.rmi.dgc.checkInterval appears to be the property that affects when a
>server object's Unreferenced method gets called. The current setting as far as
>users of RMI are concerned is 5 minutes. On a fast machine, a server PROCESS
>could have instantiated many (1,000's ?) server OBJECTS .. worst case scenario
>is the server PROCESS is configured for 1 server OBJECT per server PROCESS. To
>delay DGC (and hence regular GC .. ignoring the current bug) could cause an
>unacceptable level of resource usage for OBJECTS/PROCESS that are from the
>applications perspective dead. Imagine a Systems Manager trying to manage
>processes at the OS level.
>
>My recommendations are:
>
>1) make sun.rmi.dgc.checkInterval a java.rmi.. property
>2) allow an additional option that allows the deployer to specify that the DGC
> be additionally run after each clean call
>
>I think an alternative to 2) is to allow a client's server implementation to
>trap the clean, call the real clean in the DGC and then invoke something to
>preempt the outstanding "checkInterval".
>
>The bottom line is there needs to be more server side control available to the
>applications developer.
>
Ann Wollrath writes:
Although setting this property will only help in the case of client failure, it may be reasonable to have this as a public property so that users can configure their servers accordingly. An alternative is to make the ties between the local GC and DGC closer so that DGC is more timely.
>The sun.rmi.dgc.checkInterval appears to be the property that affects when a
>server object's Unreferenced method gets called. The current setting as far as
>users of RMI are concerned is 5 minutes. On a fast machine, a server PROCESS
>could have instantiated many (1,000's ?) server OBJECTS .. worst case scenario
>is the server PROCESS is configured for 1 server OBJECT per server PROCESS. To
>delay DGC (and hence regular GC .. ignoring the current bug) could cause an
>unacceptable level of resource usage for OBJECTS/PROCESS that are from the
>applications perspective dead. Imagine a Systems Manager trying to manage
>processes at the OS level.
>
>My recommendations are:
>
>1) make sun.rmi.dgc.checkInterval a java.rmi.. property
>2) allow an additional option that allows the deployer to specify that the DGC
> be additionally run after each clean call
>
>I think an alternative to 2) is to allow a client's server implementation to
>trap the clean, call the real clean in the DGC and then invoke something to
>preempt the outstanding "checkInterval".
>
>The bottom line is there needs to be more server side control available to the
>applications developer.
>
Ann Wollrath writes:
Although setting this property will only help in the case of client failure, it may be reasonable to have this as a public property so that users can configure their servers accordingly. An alternative is to make the ties between the local GC and DGC closer so that DGC is more timely.