确定企业Web应用程序的服务器硬件要求

Question

We have been developing a new version of our enterprise application and we are almost ready to start upgrading our customers. However since everything has changed (both back end and front end) we have to redetermine the server-side hardware requirements, more importantly number of cores and memory needed according to the number of expected users that will use the system on a daily basis.

The application back-end is written in Java and will be running on Wildfly 8.0.2 app server. Also this new version of the application is Ajax heavy and each client might be initiating 5 even more requests per second with the possibility of each request opening a database connection to fetch data.

At the moment we are not using any Java frameworks like ejb or JPA, just a simple service based 3-tiered architecture. Assuming a maximum workload of 200 users at the same time (as an example) what kind of resources this would require? As a reference, the old application (after trial and error) needed around 1 core and 2GB of ram for every 20 users I think. This time however we want to make a rough estimation before testing and fine tune it afterwards.

Answer 1

There is no rule for capacity estimation for custom software. You simply need to test it. There are plenty of tools around that can help you with that. You'll probably want to create some acceptance criteria, such as;

• ability to process X customer orders in Y minutes
• ability to serve up page Z with a response time < 200ms
• etc

Once your acceptance criteria are set the process usually goes like this;

1. Figure out what are the likely usage patterns/user journeys.
2. Write some tests that generate traffic matching your findings from 1 and also some that test the acceptance criteria.
3. Run those tests on reference hardware (just pick something; doesn't really matter what) with different numbers of concurrent users, eg 10, 100, 1000.
4. Measure how response times alter with the differing number of users.

You now have a reference point/benchmark for performance. You can say that on hardware A our software supports B number of concurrent users with a response time under C. You can expand this for different customer requirements; for example if a customer estimates double the traffic you tested with then you might want to have a server cluster to share load in order to maintain performance.

Once you have some customers in the wild using your software then you can get feedback/gather information on how it performs. This will help you refine your benchmarks.