I�d like to discuss what might be involved in making a JCS release. We
need to decide what feature should be included in a first release (or
beta release) and where I should focus.
I�m working on a few things:
1. Performance � al the bottle necks are in disk IO. I�m running
jprobe on the tester class. Since the disk gets are synchronous they
are usually slower than puts. I need to figure out a size to test at. I
usually test at 20,00 to 50,000 items in a region with only 1000 in
memory. 1 out of 20 puts is over a ms. 1 out of 6 gets all coming from
disk is 10-15 ms, the rest ar sub ms. I want to even this out.
Grouping is slower.
2. Cleaning up failover
3. Remote cache clustering. I�m not sure how the cache cluster
should work. Right now, every cluster is connected to all clustered
remote server. All puts and removals are replicated. I don�t allow
cluster gets. Clustered removal and puts are optionally sent to the
local caches. Group cache cluster may have locking problems. I want to
change the cluster to use the TCP lateral cache rather than remote. A
cluster only allows you to scale get requests and provides failover.
Deciding how cache clustering should work is more difficult than
implementing it.
4. Locking. After some thought, I�d like to add a locking
mechanism. I�m thing about adding a token ring in the remote cache
cluster. If there is no remote then lock request may return null. The
remote cache could use the lateral as the token conduit, but the lock
list. I�m not sure what to do with the lock list. . . .
I�d like to continue on the first two and consider clustering and
locking as coming features.
Aaron
