by Steve Broadhead

Table of Contents

INTRODUCTION: The Importance of Capacity Planning
Making The Case For Capacity Planning simulation tools�
NT/Sim-Exchange – An overview
NT/Sim-Exchange in use�
Summary

INTRODUCTION: The Importance of Capacity Planning

In some ways networking is like home improvements.

It’s all too easy to bolt on a bit here, glue a couple of items together there, add an extra widget here and there… the result is inevitably a short-term fix which ultimately just creates more trouble than it was originally devised to cure. And don’t think that this “tinkering” approach to network management doesn’t still go on. It does. And, yes, it does lead to potential trouble – big time.

In contrast, the real art of network management is in forward planning – ensuring that, regardless of the “component” in question, there is sufficient capacity to cope with user demands. Nowhere is this more evident than in the areas of email and messaging in general. Take Microsoft’s Exchange Server for example. What started as an unlikely alternative to Lotus Notes and even Novell’s GroupWise has acquired a huge market share of LAN-based messaging servers. Its OS platform, Windows NT and soon Windows 2000, is not exactly renowned for being entirely trouble-free when it comes to scalability issues. All of which adds up to a greater than ever need to plan ahead carefully if disaster isn’t to lie in waiting.

We’re talking about good old capacity planning - the process of determining which resources are needed to satisfy projected future requirements. While this is certainly applicable to networking and communications in general, the huge increase in the use of messaging and Exchange in particular means that it is a top priority requirement to plan email server sites in great detail. According to a report released by market analysts The META Group, there is a general increased focus on service levels and cost management� which is causing many IT organisations to investigate the potential role of capacity/performance management (C/PM) within broader IT service management strategies and email clearly is at the heart of this focus. META Group research indicates that 35% of Global 2000 IT organisations are currently addressing enterprise-wide C/PM and by 2002, approximately 75% of these organisations will have instituted dedicated teams that take accountability for end-to-end C/PM.�

META Groups’ conclusion was that effective capacity/performance management services can result in improved response time and availability service levels of critical business processes, contributing to flexibility, speed, and overall cost management. Furthermore it concluded that proactive management of IT capacity and performance requirements contributes to improved infrastructure utilisation, reduced changes, and improved application service levels.

So if we take a still rapidly growing requirement such as messaging services here, it is clear that users need to increasingly apply resources to proactive C/PM activities. By addressing C/PM requirements as early as possible within the development life cycle, META believes that the need for spending expensive resources (time, skills, and tools) during the operational stages to fix performance and capacity-related service-level problems can be substantially reduced.�

Making The Case For Capacity Planning simulation tools

A key part of the process of employing capacity planning is in bringing planning of the messaging infrastructure right down into the application planning phase and not waiting until the service has been up and running – or not – for some time.�

META, in its report, suggested that this can take the form of both stress/load testing in a prototype environment and full-scale simulation or analytical modelling. Clearly, however, this could be a very expensive exercise unless the modelling tools in question are efficient, accurate and easy to use. Which brings us to another question – should the tools be “real-time” or simulation based? There are still many individuals in networking for whom “simulation” is a dirty word. By definition simulation means “not the real world” and some take this to mean that – by their definition – it also means “inaccurate”. The belief among many is that there's far too much hype in words like "virtual" and "simulated" and that only "actual" and "real-time", when applied to networking, delivers anything that is actually meaningful.�

So – put bluntly - when it comes to offline network simulation and modelling there are still a great number of sceptics out there. Network modelling tools have been around in a minor way for years but have never become popularised in the same way that real-time tools such as network monitor and analyser products have. But there are reasons for this that need no longer apply. One of the problems historically with simulators is that they simply haven't allowed the user to model their own network accurately enough.�

The tool we’re evaluating in this report - NT/Sim-Exchange from NETfusion Inc - attempts to address this limitation. It does this by providing both the flexibility and attention to detail to model absolute elements of the network down to the nth degree and by allowing the user to either create a model from scratch or import actual data captures from their own network. The latter is obviously the closest you can get to real-time network capacity planning and evaluation without actually loading up your own network and means that effectively you get something approaching the best of both worlds.�

This is a significant point in that real-time network capacity planning and modelling tools do exist and attempt to offer approximately the same kind of features and net information as a simulation tool. Indeed we actually use such tools within our network test labs at the NSS Group. But one of the key reasons behind using such tools is that, while providing true network simulation, they also put a huge load on the network - exactly what we want in test lab conditions but hardly ideal for burdening an already busy live network with. So with real-time tools the only answer is to simulate out of office hours; that is on the few networks nowadays which don't use up all but the 24 hours a day available. This might be for either regular "user-based" work, automated backup, file updates, data consolidation, server and database synchronisation, overnight ERP runs or any other number of "round the clock" applications that now exist and, in some cases, have always existed.�

The whole essence of computers is that, unlike humans, they don't need eight hours sleep every night so to maximise their existence they increasingly get used all day and all night. For this very reason the concept of using “slack” network time to carry out real-time testing is now largely inapplicable. One alternative is to build a dedicated test network but this is equally infeasible for most companies and still doesn’t equate to the actual network you are looking to plan capacity for. Which leaves a better alternative – an accurate simulator.�

Here we’re looking to see if NT/Sim-Exchange can therefore provide the levels of accuracy and flexibility required in order to truly model a network offline – in the case for the purpose of modelling and planning a Microsoft Exchange deployment.�

NT/Sim-Exchange – An overview

An obvious – but still much asked – question when it comes to simulation tools is what do they actually simulate? In the case of a tool like NT/Sim the answer typically is anything related to impacting on network performance, service and application availability, and capacity planning thereof. But there can be a problem with this kind of modelling. Traditional simulation tools use a general approach. They analyse the performance of a process, disk or processor, but cannot typically model what causes them to perform the way they do. To do this the user needs to understand the operation of the system as a whole. Additionally, the data produced by simulation engines is generally large in volume and complex in nature, meaning that a skilled capacity planning professional is still needed at the end of the day to accurately interpret the results. Self defeating or what?

So what NT/Sim does is to take a mainframe-derived engine – the company has decades of performance management and capacity planning modelling knowledge - and effectively put a template on top of it,� – in this case Exchange Server, to do the interpretation automatically. Hence we have the NT/Sim-Exchange product. As can be easily derived from this, the Exchange module is just the first variant from NETfusion based on the generic NT/Sim engine. Others will follow soon, notably a module for Microsoft IIS web server software. Similarly, the modelling engine could be applied to other server-based applications such as unified messaging servers, ERP and CRM systems, E-business and E-Commerce hosts – basically anything which follows the format of a multi-server base.

NETfusion itself highlights what it calls application targeting as having three main benefits. First anyone who knows the application can make use of modelling techniques that would normally require extensive training in capacity planning. So the power of simulation becomes available to technical sales staff, application consultants and system administrators. Second, NT/Sim is able to construct models of a new system without the need for captured data. Providing the operator understands the expected workload, systems can be created from scratch, modelled and tested entirely in software. Third, a whole Exchange site can be processed, bot just a specific server.

In its Exchange flavour, then, NT/Sim takes the traditional capacity planning questions and answers them in Exchange-specific form. Examples include: how many servers are required in order to prevent bottleneck issues reducing service quality, what is the expected life cycle of the current NT-based network, given the planned company growth rate, or what is the likely effect of adding more users or changing the servers themselves? In a standard modelling environment the answers returned apply to the network in general and so still need a certain amount of “manual” interpretation. But with specific applications the calculations should need little in the way of interpretation, merely acknowledgement and action taken as a result of the specific recommendations the simulation throws up.�

This makes for a far more efficient – and easy to use – method of modelling than the classic, all encompassing model does. It gives a huge speed advantage when anticipating future needs in the form of and the ability to be proactive in making overall systems changes – the key element of managing a network let us reiterate.

NT/Sim-Exchange in use

The art of creating a modelling or simulation tool is to hide the complexity of the algorithmic engine behind a simple interface. It is not important for the user to understand exactly how the model is created – though they may if they wish – but merely how accurate the end results are.

In both cases here, NETfusion has scored with NT/Sim-Exchange. First, the user interface is Microsoft’s increasingly familiar MMC management console, as used for example to manage IIS web servers. Some might argue that the MMC interface is almost too simple, since, on loading up NT/Sim there appears to be very little you can do. An “NTSimEx” node appears beneath the Console Root which always sits at the top of any MMC screen and… well, that’s it. At this point you can either choose to create a modelling environment from scratch, or import data captured from a live network. Currently NETfusion directly supports Exchange server data gathered using NetIQ’s AppManager monitoring software. NETfusion provides a script to add to the NetIQ installation (simply copying a file into the correct folder) which can then be called up to capture traffic specifically for NT/Sim. This file can then be imported directly into NT/Sim-Exchange using a wizard-based importer.

Figure 1 – NetIQ Data Import Tool

There are three basic elements to NT/Sim-Exchange – Servers, User Groups and Scenarios. Within the Servers sub-node in MMC you can create the server hardware and software you wish to model. A full range of Compaq and Dell servers are currently pre-defined but you can also define your own servers, component by component, such as processor type, number and speed, amount, speed and type of memory, disk subsystem (controller, disk types and RAID levels as well as partition details) and NIC type. A wizard option provides a very easy way of creating new server entries.�

Figure 2 – The Server Wizard

Software wise, little needs to be added; what version of Exchange is being used and mailbox/folder details being the key entries. However, if you have the data to hand from traffic captured using the PerfMon network monitoring tool either it is possible to add in figures for the Information Store and Directory Cache Hit Rates. Additionally thread counts can be entered into the system for Information Store, Directory, RPC (which controls concurrent user requests) and Gateway processes. Where no figures are entered for these processes, NT/Sim uses a combination of default values and available memory in the servers being modelled to create an estimate.

User Groups describes the traffic patterns of the different levels of user group you wish to base the model around. Five groups ranging from “very light” to “heavy” users are provided as standard with usage attributes defined to match the expected usage levels. As well as the five pre-set groups it is also possible to set up a user group yourself.�

Here you can be very specific identifying, for example, what number of messages each group member generates, the distribution of sizes of those messages and where they are sent and a weighting which determines the chance of a message being received by the user in that group, rather than outside of that group. The newly created groups can then be used instead of, or along with, the pre-defined groups NT/Sim provides as standard. In this way you are able to create a model consisting of many different levels of user types, ideally for “what-if” type analysis based on what happens if a certain type of user – for example, heavy – is increased in number, rather than another type – for example, light.

Scenarios is used to create the model itself and control its execution. Here you can control what combination of servers and user groups enters the model and what kind of “what-if” changes you want to model.

So how does it work out in practice? Well, if you’re able to import data then immediately you have a system you can model. NETfusion can also provide dummy data files for you to play around with in order to familiarise yourself with the product. If designing a site from scratch there are still only a few basic input requirements. These concern specifically the number of users who are planned to be supported and what level of user they are. NETfusion recommends setting everyone as a “medium” level user initially unless you have specific traffic details to hand which provide an accurate breakdown of traffic usage per user. At this point the system can then carry out an estimate of the number and configuration of servers required to cope with the expected user demands. This is an ideal starting point for the modelling.�

Figure 3 –Initial Server Estimate

Scenarios can be created with a right-mouse click or via the top menu. Each scenario also has three sub-nodes which combine to create the complete model scenario. These are Iteration Rules, Termination Rules and Results. Using these you effectively create a modelling sequence. The Iteration Rules tell the system what to change and when, while the Termination Rules tell it when to stop, so no surprises there! New rules are created in the same way as a new scenario itself. Iteration rules are used to perform a series of change actions on either or both of Server Components and User Group properties. Typically this would be to gauge what happens if the number or percentage of certain user types increases over a period of time, or alternatively what the effect of a gradual server hardware upgrade would have on end user performance.�

Basically there are five different actions that can be selected. The Set action causes a Value to be applied to the property – such as incremental increase in users of a certain group - for all iterations of the sequence, barring an optional baseline sequence which gives a default value to work against. The Increment and Decrement actions change the property by the Value in each iteration, producing a sequence of increasing or decreasing values as the sequence progresses. Percentage versions of the Increment and Decrement options can also be used to enable the change scenario to be flexible in nature.

Termination Rules apply a limit to the sequence of actions. If just Set rules are used in the Iteration Rules then there is no need for a Terminal rule to be defined as the system will perform just a baseline and a single changed iteration. However, if Increment or Decrement rules are used then you need to set a “Terminate After n Iterations” rule, or a default of six iterations is used instead. A classic example here would be to set up the model to increase user email activity by 20% every month over a year-long period so ending up with 12 iterations of the model. It is also possible to set Termination Rules to stop when a property set in the Iteration Rules reaches a specific value, typically a traffic or utilisation threshold. Another alternative is to set the action to “Stop On Knowledge Module Suggestion” (see Issues later).�

The easiest way to create a scenario is to use the Scenario Wizard. This guides you through a series of possible actions but is designed for the “traffic increase” type modelling with the aim of determining the point at which the system fails through over-capacity, so the wizard cannot be used in every circumstance. Any models developed with the wizard can be edited and expanded afterwards however.� The details of a sequence can be edited, such as changing the results title, changing the way iteration numbers are reported and specifying the time period selected.� For example, if you are running a sequence based on predictions of the increase in traffic over your system then you may wish the report to indicate the time period you are modelling.�

There is a default time-span for each model of 60 minutes though this can be altered. In terms of the “actual” time the simulation takes to run, NETfusion advises that you can allow for approximately 50 seconds for each “job” - a job is done for each server in the site for each iteration performed – the timings being based on a PIII, 450MHz system.

The results produced come in three flavours – Graphs, Reports and Scenarios. These, in turn, are grouped into different result categories under a specific heading, such as Perceived Response Times or Issues, and exactly what mix of reports and graphs you receive depends on the nature of the simulation being run and what the outcome was. For example, Issues only appear when NT/Sim recognises that problems have arisen within your scenario which require further investigation and at this point the system might suggest alternative scenarios to run.�

The Perceived Response Time (PRT) is one of the more common results to be found.

Figure 4 – Graph Showing Response Analysis

This is derived from Microsoft’s LoadSim tool and is a “weighted average of the 95^th percentile” response time for transactions within the server. In practice it means that actions performed by the user have a greater effect on the end result than other actions. This is aimed at reflecting a users’ view of how the server is performing, so is actually meaningful! Microsoft’s own performance guidelines indicate that a figure in excess of 1000ms for PRT is too high. Bearing this in mind, then, the report highlights figures higher than this and reports them as an Issue to be investigated.

Issues typically reported are high queue times in the Exchange processes and exceeding the throughput capacity of the disk subsystems. As we mentioned, interestingly the Issues category also generates a set of scenarios designed to suggest fixes for the problems it highlights. These can then be executed simply by right-clicking on them and running the new scenario. This will produce results just like any other scenario and can raise further Issues and ensuing optional scenarios to run

Figure 5 –Issue Raised And New Scenario Suggested

Another possible result the system might generate is the System Queue Point Analysis, typically only generated if there is a problem the scenario has highlighted. This can be activated through the Scenario Results Selector page of the scenario properties. One report is produced for each server per iteration in the sequence.� It shows the activity caused by each Exchange process, including the simulated RPC process representing the network activity generated by user requests. The report shows the amount of processing and disk activity generated. If this activity cannot be completed within the time period of the model, the activity will be marked as over-utilised and generates an Issue.

Probable Queue Length is also reported for each aspect of each process. Three figures are quoted: an average figure (50%), a figure covering most cases (95%) and a worst case scenario figure (99.99%). High numbers here tend to cause slower server response, especially when they apply to the Store or Directory processes so these will often create Issues. Transaction Response Analysis is another results screen which is typically only seen in the event of a problem being identified by the modelling. These reports are provided along with a bar graph which shows the time taken from transactions to pass through the server, Processor Activity columns report the response considering only the activity within the server while the Response Time columns also consider network activity.��

Figure 6 – Transaction Response Analysis Report

Each row in a report indicates the time taken to pass through each stage of the transaction. Large jumps in the time indicate a delay due to activity in a process. There is one transaction report for each type of transaction within the server, each of which has a frequency against it, indicating the number of these transactions generated by the model. At the end of the report is a summary, including an arithmetic average response for all transactions within the server, the total number of occurrences for all transactions and the perceived response time statistic for this server and will be highlighted if it exceeds one second.

Summing up NT/Sim-Exchange in use, anyone familiar with MMC will know immediately how to use the “mechanics” of the system. For those who are unfamiliar, MMC is like Windows Explorer in operation, consisting of a left-hand window – the tree view - displaying the hierarchical structure of the “system”, in this case, NT/Sim. So each of the three sub-node categories, Servers, User Groups and Scenarios, can be expanded by clicking on a plus sign attached to each and likewise for the sub-nodes which appear beneath.�

In addition there is an Action drop down menu at the top of the screen which displays, at any time, a list of actions that can be performed on the object highlighted in the tree view. Better still, with NT/Sim you can just use the right mouse button to access the same menu options, so operation is very slick indeed.

We modelled several different types of scenario, including many variations on the “company expansion” model as well as merger scenarios and even server consolidation. We found in all cases that the simulation produced meaningful, easily understood results with problem areas clearly highlighted in the reports. One thing the product makes easy is trend analysis.

Figure 7 –Trend Analysis Of Different Server Components

This is not simply trend analysis of a given element of the model over time – which NT/Sim does extremely well – but also comparing the behaviour of one component against another over a given period of time. So it is very easy to see where the problems lie within the server or servers down to the specific component within a given server.

As tested on an HP Vectra PII/450 with 128MB RAM, running Windows 98, and depending on the extent of the model scenario, the system took anywhere between a couple of minutes and several hours – literally – to do the number crunching and create the results. At this point it may be necessary to run one of the suggested alternative scenarios so total run-time can be extensive. But the great thing here is that it runs off-line on a standalone PC and therefore does not impact on the network at all, of course. You can even take your modelling home with you in the evening on the laptop and let it run through the night.

The final point to make here concerns the issue of accuracy, the second vital ingredient of a successful simulation tool. Not surprisingly NETfusion has made great efforts to ensure and test the accuracy of NT/Sim-Exchange, not least with the benefit of 30 years of primarily mainframe capacity planning and performance management simulation behind the company.�

It has also tested its own simulations against real-time benchmarks, notably comparing some extensive benchmarking Compaq carried out using Microsoft’s LoadSim tool on its own servers running MS Exchange 5.5, with the equivalent NT/Sim-Exchange simulation. As you can see from the graph, the NT/Sim results were remarkably close to the real-time performance results taken by Compaq, which is a great argument in favour of simulation. In order to carry out its benchmarking, Compaq had to run on a live network. In the case of the NT/Sim model, this was carried out on a standalone PC.

Meantime NETfusion is continuing to test its own models against those captured in a real time environment, in an attempt to continue to fine-tune, and thereby increase further the accuracy of, its simulation engine. The company is also looking to extend its server database significantly and add in support for newer technologies such as SANs within the disk subsystem categories, currently restricted to mainstream SCSI/RAID combinations. Overall, however, we have no reason to doubt the accuracy of the models and NETfusion appears to be doing all it can to ensure that this level of accuracy is maintained or improved in the future.

Summary

Overall, NETfusion seems to be heading in a very interesting direction with NT/Sim-Exchange. Historically, too many modelling and simulation tools have been vague – both in operation and results – but NT/Sim, by taking a specific application to model, is very direct and very usable as a result. It scores on flexibility too, in every sense of the word. One example of this is the regular option of either using wizards to create a quick but relatively limited element of the model, or to completely customise it – and this is not complex – and create a very accurate scenario to model. There are some limitations, such as the initial server estimate which restricts itself to groups of 1,000 users maximum, but we understand this issue is being resolved anyway.

What NETfusion needs to do now is to build up a library of application support for the NT/Sim engine and we know that this is indeed what is happening. First up is Microsoft’s ISS web serve but other applications will follow on from this. In practice, with the simulation engine in place, the NT/Sim approach could be applied to any server-based application where scalability and capacity planning are the issues for multi-server deployment.

Send mail to webmaster with questions or�
comments about this web site.
Copyright � 1991-2002 The NSS Group.
All rights reserved.