GSS architecture

Handling a large number of concurrent connections requires many servers. Not only because scaling vertically (throwing bigger hardware at the problem) is very costly, but also because even if an application can be designed to scale vertically, the underlying stack probably can not. Java applications for instance, like GSS, run on the JVM and although the latter is an excellent piece of engineering, using huge amounts of heap is not something it's tuned for. Big Iron servers with many cores and 20+ GB of RAM are usually running more than one JVM, since garbage collection is not all that efficient with huge heaps. And since running application instances with a 4-8 GB heap size can be done with cheap off-the-shelf hardware, why spend big bucks on Big Iron?

So having a large number of servers is a sane choice, but brings it's own set of problems. Unless one partitions users to servers (having all requests a particular user makes be delivered to the same server), all servers must have a consistent view of the system data, in order to deliver meaningful results. Assigning user requests to particular servers, usually requires expensive application layer load-balancers or customized application code on each server, so it would rarely be your first option. Having all servers work on the same data is a more tractable problem, since it can be solved by having the application state being replicated among server nodes. Usually, only a small part of the application state needs to be replicated, for each user, that is the part which concerns his current session. But even though session clustering solutions have been a well studied field and implementations abound, having no session to replicate is an even better option.

For GSS we have implemented a stateless architecture for the core server, that should provide us with good scalability in a very cost-effective manner. The most important part in this architecture is the REST-like API that moves part of the responsibility for session state maintenance to the client applications, effectively distributing the system load to more systems than the available server pool. Furthermore, client requests can be authenticated without requiring an SSL/TLS transport layer (even though it can be used if extra privacy is required), which would entail higher load on the application servers or require expensive load balancers. In the server side, API requests are being handled by servlets that enlist the services of stateless session beans, for easier transaction management. Our persistence story so far is JPA with a DBMS backing, plus high-speed SAN storage for the file contents. If or when this becomes a bottleneck, we have various contingency plans, depending on the actual characteristics of the load that will be observed.


The above image depicts the path user requests will travel along, while various nodes interact in order to serve them. A key element in this diagram is the number of different servers that can be found, effectively specializing in their own particular domain. Although the system can be deployed on a single physical server (and regularly is, for development and testing), consisting of a number of standalone sub-services instead of a big monolithic service is a boon to scalability.

This high-level overview of the GSS architecture should help those interested to find their way around the open-source codebase and explain some of the design decisions. But the most interesting part from a user's point of view would be the REST-like API, that allows one to take advantage of the service for scratching his own itch.

So that will be the subject of my next post.