Hello fellows,

I write to you in search of an advice. I have a decision to make and would appreciate some considerations.
We have a product developed as a web application. The performance of the software is good, and it is able to handle many users. We don't have problems neither with memory and neither with database performance (At least so far).

The problem I am facing is with connection bandwich and latency. Our biggest client has offices in a lot of countries. Currently, some of the stations have really slow connections. Our interface is pure javascript (somewhat similar to gwt) and a lot of javascript is needed to download. Even with caching, it is slow to access, due to the very slow connection. What I would like to do, is to clusterize our application, to provide, in the slow nodes of the client, a server which would provide a fast, nearly local connection.

The architecture of our application is rather "simple".
Our application consists of DWR + Google Guice 2 + Hibernate 3.2 + Hibernate Search. We distribute it together with jetty 6, rather than installing it on client's web container.

Almost all of our services have a public interface, which is visible for dwr to call, a private interface, which is used to connect different services, and a implementation class which implements both interfaces. Guice handles the dependency injection on these services.

We have, in some points, objects which are held in the memory. These are information which almost never change, and when a writing operation is needed, we write both in the entity and on the memory. These consists of country information, organizational structure information and permission profiles. We have a really complex permission algorithm, thats why we keep them in memory.

My question to you is: How can I clusterize this scenario? Which approach would you take?
Please consider in your answer that my biggest concern is time + human resources (=money).

If you need more information, please do ask.

Thanks for your time,
Guilherme

you could try keeping as less as possible in memory and have an additional server "close" to the clients but still using the same database.
This would mean that accessing the database would be quite slow, as the more close you are to the clients with slow connections, the farther you are from the database, but at least you could have a connection pool to the database with ready-to-use connections. I'd recommend a good connection pool as you will need to monitor these connection's health as well.
You could then use a clustered second level cache like Infinispan to aggressively cache most state from the database, and reuse infinispan to keep the complex security state you mentioned as well, so that it gets properly invalidated to each server when you change some permission.
It's a trade-off however, as if your code will trigger many network messages and your network is slow, this might get quite slow.

Using Hibernate Search cleverly for most frequently run queries will definitely help performance, as the searches are performed locally and won't need to run a query on the database; with projections ans second-level cache loading (Hibernate Search 3.4.0) you can retrieve full results even without a database connection.

it's not easy, but can be done. not sure how cheap you can make it, that really depends on your skills. switching to JBoss AS from jetty might help to manage and monitor the connections health and timings, and clustered caching.