The First Step
It is easy to agree with the need for a performance model, but how do you build one. Here is one approach that is suited for a transactional site:
You have just decided to replatform your general merchandise eCommerce site. What should be your first step?
A lot of companies start by contacting software vendors and build partners , getting quotes, writing detailed requirements, creating new business flows, etc. While all of these are needed, they miss the most important step, the performance model. A performance model is simply the number of transactions that must be accommodated, detailed by type, along with the expected response time for each transaction type. For example, a Home page view is one type, taking an order is another, taking a return order is a third type, etc. A performance model must be one of the first steps, if not the very first, because without it, everyone who will be involved in engineering the site will be guessing at the size of the task at hand.
Start with the orders/day. Most companies know how many orders they take every day. Next, you plot the trend line of orders in 2007 verses orders last year verses orders this year. This will give you the year over year growth trend. Using this line, you can create a rough estimate of next year’s and future year’s traffic. Then factor in anything that will make next year and future years different than the past. If you are in a recession, you may flatten the trend line some. If you are investing a lot of money in advertising the site, then you will steepen the line.
Next, translate orders into that into page views. The best source of this information is from your Web analytics vendor like Coremetrics or Omniture. Your own Web server statistics may provide these also. Armed with this, you will be able to create a ratio of page views/orders. For example, if you take 2000 orders/day, and serve 2,000,000 page views/day then your ratio would be 1000:1, which means that on average, every time your customers view 1000 pages, they place an order.
Following this, you can project the number of page views/day based on the order data. A complete model requires additional data like “number of returns” number of order status queries, but those too can be estimated using the historical ratios of returns/order and queries/order. Once you get the business to bless these numbers, you can contact the vendors of all stripes and provide them with consistent load data from day one.
It is critical to publish this data to everyone involved in the project in order to avoid having multiple estimates floating around. I have seen projects where each development team was given a different load model.
