This post originated from an RSS feed registered with Java Buzz
by Talip Ozturk.
Original Post: JPower 1.0 is out
Feed Title: Shared Memory
Feed URL: http://sedoparking.com/search/registrar.php?domain=®istrar=sedopark
Feed Description: about java, jini, javaspaces, programming, aop and a little bit of me and life.
JPower1.0 is available for download. So what is next? Biggest next thing is the clustering of course and I have been working on it for about a month now. I have got load balancing and fail-over implemented. I have got to tell you, clustering is so cool. You start up couple of JPower instances, you shutdown one or two, failed invocations will fail-over to the remaining available servers. You start the dead servers back up, they join the cluster right away. It balances the load based on selected policy. I only implemented round-robin though. but implementing a others is not a big deal. I will implement random-robin, and weighted-robin soon. these are not so much fun for me though; real fun will be implementing a policy that has AI in it. Imagine this, an AI based load balancing policy, which is downloaded by client a long with proxy, takes client's first ,say, 1000 invocations as its training data and analizes them then determines which cluster node is the best for which type of invocation. It may end up routing -write- invocations to one server, -read- invocations to another and -transactionals- to another.
These are all good but I want to mention one important thing. Here is the story... About 2 weeks ago, I wrote about life cycle of cluster aware proxy. In JPower, I was willing to dynamicly update the information on the proxy that client downloads so that when one server goes down or comes back up, the proxy will know about it ASAP. One cheesy solution to have proxy to ping the dead services everytime. This is obviously too costly. Second bad solution was to have proxy to ping dead services ,say, after every 100 invocations. That is less costly but at the same time less accurate. I want proxy to know that a previously dead server is back up and ready ASAP. Markus Kling sent me an email and suggested a solution. Here is the solution [I rephrased it in my words.]
keep the proxy stateless. decouple the cluster info.
proxy passes its version to the server along with the invocation. if the version on the server different new cluster-node list goes to proxy a long with the invocation response.
share the cluster info among the proxies in the same wm. that eliminates the fetching cluster-node list from server again for the other proxies in the same jvm. all the proxies in the same jvm in-synch.
I really liked what he suggested and I implemented it. This is good stuff, I mean you put up a question, you write down your concerns, someone you don't even know returns you with answers. It doesn't really matter if it is the best answer or not, someone else tomorrow returns you with even better one. Blogging is better and more meaningful if people are humble/kind enough to return you with their opinions and guidance. Thanks Markus.
Previous Topic
