Close Help Window



Overview This model illustrates the relationship between various demand curves, capacity strategies addressing those types of demand singly or in combination, and the economics of tradeoffs. Is it better to use solely static capacity engineered to peak demand? Or to use less static capacity and accept penalties associated with inability to meet peak requirements (e.g., loss of revenue associated with unserved customers)? Or use purely utility resources to meet all demand? Or to use a hybrid approach? The answer is: it depends, and this model shows the relationships between the different factors that would lead to an optimal selection.
Basic Flow First, demand is characterized using the Demand Shaping window. It may be uniform, worker, gamer, event, cyclical, or growth -- or a combination of some or all of these. For example, worker demand occurs from 9:00 AM to 5:00 PM, Monday through Friday. In any of the hour time slots, the demand falls between the minimum and maximum number of servers.

Second, the resulting demand curve may be viewed in either the Annual or Weekly Demand windows. The Weekly Demand window is a bar chart of every single hour of a one week period. The Annual Demand window shows the maximum, minimum, and or average hourly demand for each of the 365 days in the year.

Third, a variety of capacity strategies may be tried. By allowing additions and reductions to capacity, it may be possible to reduce the expenses associated with fixed capacity. The challenge is that too little capacity may lead to some demand not being served, which has a financial penalty associated with it. And, changing capacity may incur charges.

Fourth, cost assumptions may be modified. Whether costs for fixed capacity, costs for capacity modifications (installs or decommissioning), opportunity costs of unserved demand, or costs for utility servers.

Fifth, statistics may be viewed, e.g., minimum demand, average demand, unserved demand hours.

Lastly, the relative costs and underlying drivers of four different types of architectures / business models may be viewed as a bar chart.
Characterizing Demand
The demand shaping window is used to characterize the nature of the demand. Each demand type has different parameters that may be used to shape the curve. For example, the cyclical demand type has controls to adjust amplitude (maximum number of servers), period (time it takes for the cycle to repeat itself), and offset (shifting the peak of the cycle forward or backward in time). Checkboxes indicate whether the demand is to be included in the analysis or not. Finally, the smoothing parameter at the top of the window may be used to smooth demand. It specifies the maximum change (delta) in any time period. Lower numbers smooth demand, whereas higher numbers may leave it more jittery.

Uniform Demand -- Uniformly distributed between min servers and max servers.

Worker Demand -- 0 after hours, but uniformly distributed weekdays from 9 to 5.

Gamer Demand -- Gamers are assumed to have day jobs, but are active nights and weekends.

Event Demand -- Used to model special events, e.g., concerts, disasters, movie premieres. The parameters specify the typical duration of the events, the typical time between the end of an event and the start of the next, and of course the minimum and maximum demand. "Typical" times are averages, and durations may vary from 50% to 150% of the specified duration.

Cyclical Demand -- For daily, weekly, monthly, quarterly, or annual cycles.

Growth Demand -- Starts the year within a particular min / max range, and is smoothly interpolated to a different min / max range. This can also be used to model declines, or even a change in variance (e.g., the year may start with demand in the range from 0 to 100, but end with a more compressed range, say, 40 to 60.

Weekly and Annual Demand
The annual and weekly demand may be viewed. Clicking on a daily demand bar or entering a week number changes the week that is "zoomed in." Hovering over a demand bar shows the time slot and information associated with it. Hovering over the blue capacity graph shows the capacity level and the duration it is in effect.
Capacity Strategies
A variety of capacity strategies may be attempted. The initial capacity and minimum capacity may be adjusted. Capacity increases may be allowed, based on recent peak, average, or minimum demand values. Delay times for placing the capacity "order" may be adjusted. Separately, capacity decreases may be similarly allowed. However, there are costs associated with having capacity (in effect, monthly lease costs or depreciation), capacity increases (installation and turn-up), or capacity decreases (e.g., transportation, removal, reconditioning). Different types of demand curves may be more or less amenable to different capacity strategies. Flat demand, with little variance (min close to max) can effectively be supported by a fixed capacity strategy. Steady growth can be amenable to relatively infrequent upgrades which may be planned in advance. Cyclical requirements can be met by alternating increases and decreases. However, trying to rapidly respond to random demand changes, like buying stocks after an up day in the market and selling them after a down day can be counterproductive. If response time is too slow, capacity may be being decreased after demand has begun to increase. And, trying to follow every twist and turn in the demand curve can cause churning of the capacity, incurring excessive charges.
Cost Assumptions Cost assumptions are key to deriving useful results from the model. Increasing or decreasing capacity in an enterprise data center is not likely to be free, nor instantaneous. While it is relatively rapid in a utility environment, such as a service provider cloud, it should be assumed to be more expensive than fixed capacity, on a per hour basis. This matches real world premiums typically associated with a utility. For example, a rental car costs more per day than owning a car (per day charges). A service provider can create a shared utility by statistically multiplexing multiple demand sources onto a resource, but whatever that resource costs is just the baseline. Also added in must be a capacity allocation manager, infrastructure upgrades, margin, customer acquisition costs, utilization "breakage," and so forth. Consequently, the optimization challenge is to figure out the right balance between fixed resources, which cost money whether they are used or not, and utility resources, which cost more when they are used, but cost zero when not used. Another important cost assumption is the opportunity cost of unserved demand. If it costs nothing to have angry customers for whom there was no service capacity, then it is cheapest to minimize fixed capacity. A key assumption must be that each unit-hour of fulfilled demand creates revenue or other value well beyond the cost to ensure the availability of capacity to so serve that demand.
Statistics
A variety of statistics are automatically generated each time any assumptions or inputs change. These include things like the maximum demand, the minimum demand, the average demand, and how many unit-hours of demand would be unserved due to capacity shortfalls. Also, total charges are displayed for four different models.

A pure pro forma fixed capacity model. This assumes unvarying capacity as high as peak demand.

A variable capacity model, with either capacity increases, decreases or both. Deltas may save capacity costs, but also run the danger of creating additional cost through unserved demand.

A pure utility model. The capacity model (fixed or variable) is ignored, and unbounded resources are assumed to be available in the cloud, but a premium is paid for each unit-hour of capacity.

A hybrid model. Capacity as specified is used and charges incurred, AND utility resources are leveraged to ensure that there is no unserved demand.
Total Charge Visual Comparison
The pure pro forma fixed capacity model, the variable capacity model, the pure utility model, and the hybrid model are displayed graphically so that they may be compared. Adjusting capacity strategies, demand curves, and/or cost assumptions alter the relative costs.
About This model was written by Joe Weinman in just over 2000 lines of code, which are a mixture of HTML, DHTML, ASP.NET 2.0 / Visual Basic, stylesheets, and JavaScript, using Microsoft Visual Web Developer 2008. The ASP.NET AJAX Control Toolkit Extensions almost worked successfully. It has been tested in Firefox 2.0 (Mozilla 5), Safari 3.1, and Internet Explorer 7.0.
 
 
 
© 2005-2008 Joe Weinman