Driving by the Rear-View Mirror:
Managing a Network with Cricket

Jeff R. Allen
WebTV Networks, Inc.

Note: This paper was first published in the proceedings of the 1st Conference on Network Administration, which was held April 6-10, 1999. This paper is copyright 1999 by Jeff Allen. Redistribution of this paper is allowed under the GPL.


Cricket is a tool that lets users visualize a set of measurements over time. It was designed to assist network administrators by letting them see and respond to patterns in their network. In this paper, I will describe the need we saw and attempted to resolve by writing Cricket, then describe the solution we came up with. Finally, I will describe some future work we expect to do to make Cricket a more proactive monitoring tool.

The Need

When running a complicated network, there are a multitude of things one needs to keep an eye on. Clearly, immediate concerns like connectivity, link status, and routing stability are top in most administrator's minds. Long-term issues like architecture and technology decisions float in the back of our minds. Often, that leaves no room for the most interesting questions, which help with both short term issues and long term ones. Some of the questions we found ourselves asking about various network components were:

Taking the time to ask these questions allows us to step up a level in our monitoring, and move from a reactive realm (i.e. Is the link to Europe up?) to a more contemplative and predictive realm (i.e. Are we seeing an unexpected burst of traffic to Europe?) It's obviously not possible to definitively say that this makes our job easier and our network run more smoothly, but it certainly seems that way. Simply having the data available to be able to say ``things look OK'' gives us a peace of mind that simply reacting to the network would never give us.

Figure 1 shows a screenshot from Cricket, showing the normal way that data is displayed about a target. In this case, we are looking at the traffic on one of our OC3's, which was operating normally until an outage started a little after noon. From this same view, we can tell that in the last week, the peak bandwidth in use on this link was 60 Mb/sec, on Wednesday evening. By glancing at one page, we can answer many of the questions posed above. By looking at longer time scales (accessed via the links in the upper right) we could answer still other questions about the long-term behavior of the traffic on this link.

Figure 1. An example of Cricket in action. These graphs show an OC3 outage.

One of the reasons it's hard to monitor a network is that there are a lot of different components, each with different operating characteristics and monitoring needs. For instance, in our network, some of the components that we use Cricket to monitor are:

Note that we monitor some things (DNS, hosts) that are not directly related to the health of the physical network. In our shop, the ``host folks'' and the ``network folks'' work together very closely. There is no artificial separation between the two. I will focus on monitoring the networking components, however, keep in mind that Cricket is capable of monitoring many types of components.

Just as the components we wish to monitor are varied, so are the ways in which we talk to them. Sometimes, we can simply talk SNMP, then either record the data directly or do a bit of post-processing to derive a rate of some kind. Other times, we fetch data via SNMP, then post-process it in some more complicated way to get a final data point. For instance, to monitor modem bank usage, we fetch the current state of all the modems, and count the number that are off-hook. Sometimes we simply run a shell command on another data-gathering system. For instance, we can use Unix tools like 'wc' or simple Perl scripts to derive a data point from the data already collected by syslogd. Finally, sometimes we want to measure and observe an aggregate of other data sources. In a site with multiple Internet links, it would be interesting, for instance, to plot the total Internet bandwidth across all links.

MRTG to the Rescue

Around the time we1 were bringing up the WebTV network, we fetched and installed a tool called MRTG2 to let us start answering the kinds of questions mentioned above about our WAN links. In fact, MRTG was in use at WebTV Networks before we had brought in any commercial network management tools at all. It was only by using, and coming to depend on the MRTG graphs, that we were able to take a step back and figure out why they were so useful to us. After getting ourselves thoroughly addicted to MRTG, even while attempting to roll out another network management system, we were forced to ask the question, ``What is it that MRTG does that we can't get elsewhere.'' The answer to that question makes up the bulk of the first section: MRTG let us manage our network better than reactive systems.

So, there we were, addicted to MRTG, using it in new situations and in ways it was never designed to be used. MRTG started showing signs that it would not scale to handle the new jobs we wanted to throw at it. Something needed to be done, and quickly. A true addict will not wait patiently for a new drug!

Before we threw out MRTG and started over, we thought it might be a good idea to figure out what it does right, and make certain that whatever we got to replace MRTG could at least do those things. Each view of the data has just enough detail to tell the story, but not enough to hide the good bits. For instance, the weekly views show today and this day last week. There is an enforced data density so that graphs have just enough data to tell an accurate story, but not enough to overwhelm the viewer. In addition, the auto-scaling feature consistently produces readable graphs (though there is a bit of room for improvement on this front). It uses a fixed size database which grows linearly with the number of monitored devices, not time. It needs no extra ``cleanup'' work. Because MRTG is web-based, we have a platform independent tool: all it takes is a web browser to check the status of the network. You can even use a WebTV®-based Internet Unit to see graphs of your network on your home television!

Perhaps more important than all that, though, MRTG passes the ``can it solve real problems'' test:

To see how we use graphs to identify and understand real problems, examine figures 2 and 3. These two graphs are selected snapshots of the graphs available to us during backbone routing instability on one of our peer's networks. During this event, it's clear that about 10 Mb/sec of traffic shifted over to the second peer. We were able to use Cricket to find the nature of the problem, see that the outbound traffic had failed over to non-optimal, but functional link, and finally verify that the traffic returned to normal after the event.

Figure 2. Traffic on a Fast Ethernet handoff to a peer of ours. Their backbone was suffering routing instability at the time.

Figure 3. Traffic on a DS3 to another peer during the routing instability shown in Figure 2.

Finally, Cricket passes the ``manager test'' with flying colors. Managers immediately understand the need for a bandwidth upgrade once they see a graph produced with MRTG. Even more astoundingly, it even passes the ``executive test'' which is very important since executives tend to need to sign orders for OC3's and above. The simple web-based user interface makes it easy to print and share the graphs with management, making it easier for everyone to do their job.

Regardless of all these great features, we identified problems we wanted to address in the next generation system. First, MRTG has a narrow view of the world: all targets have exactly two data sources (used by MRTG for bandwidth in and bandwidth out). We wanted to be able to measure other things about components. MRTG also has severe performance problems, resulting from the way it reads and writes data. MRTG was incrementally developed and the performance problem only became apparent after it was too late to fix it. Usually, it's not an issue, because MRTG is mostly used in small installations. However, with all the different ways we were trying to use MRTG, performance became a problem for us.

There are ways to overcome these performance problems by arranging to run multiple instances of MRTG in parallel. However, the configuration needed to run it in parallel is even more complicated than the usual configuration. For our varied uses, the configuration was already unwieldy and inefficient. The last thing we needed was to make it even more complicated!

Why Not Commercial Tools?

Some might recommend commercial products as a replacement for MRTG, since commercial products are supposed to be designed to scale well and be easy to configure. Our experience is that commercial tools introduce as many problems as they purport to solve, and often leave the original problems (scalable performance and configuration) unsolved. Here's an undoubtedly one-sided list of the problems inherent in most commercial tools that could have helped us.

First and foremost is cost. It's expensive to make good software. No one will argue that point. In fact, it has been more expensive than we expected to develop Cricket. But when buying an expensive network management tool, things get ugly. It takes too much time, paperwork, patience, and persuading management, even if you do have the budget for a commercial tool. Evaluating and rolling out a free software package can happen in an afternoon, if you put your mind to it. There's simply no downside risk - install it and if it solves your problem, keep it. If it doesn't meet your expectations, delete it and move on. You get what you pay for, but sometimes it's what you are not getting (support hassles and licensing nightmares) that really counts.

With commercial tools, database formats are often secret, API's (usually only available at extra cost) are either a pain to work with, or are major performance pigs. Customer support varies, but the simple fact is that even stellar customer support won't do for you what getting your own two hands dirty in the code will do: give you a deep understanding of your NMS and the confidence to trust the system and your ability to operate it. Then, there's the version tango. The NMS doesn't find out about new router models until 6 months after you've already deployed them because the router company and the NMS company are feuding. The trouble-tracking system you are using is only certified to talk to the next version of the NMS that you cannot use yet because that version introduces a critical bug that no other customers are seeing, and thus, will not get repaired until you threaten to revoke your support license - which, by the way, does not expire until next December.

Finally, let's be honest: Free Software rules. Who would you rather trust with the smooth operation of your network, a bunch of folks who are in the same boat as you, or a giant vendor with 100-page implementation plans and slick salesmen? The choice is obvious, to me at least: network managers design better tools to do their job3. Sometimes commercial tools are the right tool for the job, but in this case, it made more sense to go at it on our own. It will make even more sense for your organization, since WebTV Networks has done much of the hard work. All you need to do is download, install and configure Cricket.

The Solution

The solution comes as two pieces, a new database and a new user-visible front-end. By changing out the back-end, we addressed the flexibility and performance goals. By changing out the front-end, we improved the user interface, both for configuring the system and viewing the graphs. The two pieces are developed semi-independently, one by me (in California) and one by the original author of MRTG, Tobias Oetiker (in Switzerland). Keeping the two pieces separate makes it easier to develop them, and it also means the database component can be reused in other contexts. RRD, the database backend, is distributed separately from Cricket4.

We are deeply indebted to Tobias for his hard work on the database back-end. Without it, Cricket would be nothing. By engineering RRD correctly from the start, and by improving it's reliability over the last year, Tobias solved over half our problems with MRTG himself. I simply had the easy job of passing the right data into his code. All the seeming magic of Cricket comes from Tobias's RRD code.

RRD: The Round Robin Database

The next generation back-end is called RRD, the Round Robin Database. It takes over exactly the same jobs the back end provided with MRTG did: storing and rolling up the data and generating graphs from the stored data. RRD is written in C, and comes in both a Perl module and a command line version, which can be used interactively or across a pipe from scripts. It can be used either directly by simple scripts, or via frontends like Cricket. Other frontends are available from the RRD website.

RRD achieves high performance by using binary files to minimize I/O during the common update operation. The ``round robin'' in RRD refers to the fact that RRD uses circular buffers to minimize I/O. RRD is at least one, and possibly two orders of magnitude faster than MRTG's backend, depending on how you measure it. It's truly incredible to watch it chew through data if you have ever seen MRTG trying to handle the same job.

RRD is more flexible than MRTG in at least two dimensions. First, it can take data from an arbitrary number of data sources. MRTG was limited to two datasources. Originally they were reserved for input bandwidth and output bandwidth, though they got co-opted for other measurements by many MRTG hackers. RRD can also keep an arbitrary number of data arrays, each fed at a different rate. For instance, you can keep 600 samples taken every 5 minutes alongside 600 samples taken every 30 minutes. Thus, you can have 5 minute data stretching back 50 hours into the past, and 30 minute data stretching back 12 days. When you draw a graph of this data, you can see recent data in high resolution, and older data in lower resolution. This is one of the original features of MRTG that made us so happy. In RRD this feature is completely configurable.

RRD stores data with higher precision than ever before (float versus integer) which allows us to measure bigger things (OC3's) and smaller things (Unix load averages) without relying on goofy scaling hacks, as we were required to do with MRTG. The data file where this data lives on disk is still a fixed size over time, and scales with the product of the number of datasources and number and sizes of the data arrays. The data storage in use for an entire installation still scales linearly with the number of targets under observation. At WebTV Networks, we have Cricket configured to store 6 variables for every router and switch interface. Each interface requires 174 kilobytes on disk, which is enough room to store all the data for 600 days. Of course, your mileage may vary; Cricket is configurable, so you can choose to keep more or less data, depending on your needs.

RRD still provides the same simple, sparse graphs we grew to love while using MRTG. There is now much more flexibility at the time the graph is generated. For instance, we can choose to show some data sources, but not others. We can choose to scale a data source using an arbitrary mathematical expression. For instance, we can fetch the temperature from a router in Celsius, and present it to the user in Fahrenheit. Since almost all data seems to arrive in exactly the wrong units, it's quite helpful to have this flexibility. Do you think about Ethernet capacity in ``megabytes/sec'' The scaling feature lets us turn that measurement into ``megabits/sec'' It is also possible to integrate data from multiple sources into a single graph. For instance, you could make a summary graph showing the sum of the traffic across all of your Internet links.

Figure 4. A simplified example of a config tree.

The Config Tree

Recall that part of the problem was flexibility and performance of the database, both of which RRD solved. The other part of the problem was scalability of the configuration. Our solution is something called the config tree. A config tree is a set of configuration files organized into a tree. This hierarchical configuration structure allows us to use inheritance to reduce repeated information in the configuration. To simplify implementation of the config tree, the inheritance strictly follows the directory structure; complicated concepts like multiple inheritance are not supported. The attributes that are present in a leaf node are the union of all the attributes in all the nodes on a path leading from the root of the config tree to the leaf node. Lower nodes can override higher nodes, so it's possible to make exceptions in certain subtrees, and do other clever sleight of hand.

It's easier to understand the config tree by looking at an example (see figure 4). Attributes that all of the system will share are located at the root of the config tree. For instance, the length of the polling interval is set there. At the next level, we set attributes that will be restricted to the current subtree. At this level, typically you will find the target type. Finally at the lowest level we set things that will vary on a per-target basis. For instance, we set the interface name that we are trying to measure here. By using the power of inheritance, you can avoid repeating the information at the top of the config tree many times near the bottom of the config tree. The three level config tree in the example is the simplest in common use. The one that ships with Cricket in the sample-config directory works like this. Config trees can and do have many more levels. At WebTV Networks, for instance, we add levels to break apart routers in different data centers to make the directory listings more manageable. There are no built-in limits on the shape of the config tree, only practical ones.

Figure 5 shows the power of the string expansion feature in Cricket. When it comes across strings in the config tree with special markers in them, it expands the strings in much the same way a scripting language expands variables. The sample config tree that ships with Cricket uses this feature in several places to dynamically build strings from settings already available. In the example, the short and long descriptions for the target are set based on some data inherited from high in the tree, and other data related to the target itself. In order to make this feature even more useful, we introduced auto-variables, which are set to certain strings automatically by Cricket. They are available for use by expansions, making it possible to automatically generate things that change for every target, like the data file location. The same expansion system is currently used to make it possible to customize the HTML output of the user interface component, though it is still not yet as flexible as we would like it to be.

Figure 5. An example of variable expansion.

Cricket's collector is single-threaded, and thus spends a fair amount of time waiting for data to arrive over the network. The wall-clock run time for each instance of the collector is limited to the length of the desired polling interval (or else you might miss polling cycles, which would be unfortunate). Thus, the number of targets that can be polled by a single collector on a single host is limited. To get maximum polling capacity, it is necessary to run several collectors in parallel, each working on a different subset of the total set of targets to be polled. This was hardly a surprise to us, since we needed to use the same technique to boost the performance of our MRTG installation. In fact, we designed the config tree to make partitioning the set of targets easy, so parallelizing Cricket is simply a matter of starting the right collectors at the right time, each operating on the right subtrees of the global config tree. There is a wrapper script for the collector which does this, as well as other housekeeping chores like managing logfiles, and checking them for errors.

The config tree has several other advantages we stumbled upon after using it for a while. Because the subtrees are mostly independent from one another, multiple administrators can work on the config tree without impacting each other. The hierarchical structure makes it simple to make standards that apply across all the different applications. It also makes it possible, when necessary, to make exceptions for certain devices. Of course, in a perfect world, this wouldn't be necessary. But this isn't a tool for a perfect world, it's a tool that was written to solve problems in the real world. Sometimes it just happens that we need to query one device with different OID's, or that the community name is different on a certain device. At the same time, though, a properly designed config tree can save you a lot of work when you decide to make some kind of wide-spread change to the system. Instead of changing the data everywhere, you can change it in one place.

Contemplating our fully populated config tree gave us another idea. One of the hardest parts of maintaining a suite of network management products is trying to keep their configurations up to date and in sync with one another. What if we declared, by fiat, that our Cricket config tree was the one true source of configuration information, and that all others should be derived from Cricket's config tree? We decided to design the config tree to be a simple collection of dictionaries. As the config tree is read the parser pays no attention to what keys and values it finds. It simply stores them away. Later, the application which called the parser (Cricket in this case) does lookups in the dictionary looking for data it understands which will control its behavior. Since it is doing lookups into the data, it never even sees data that it's not interested in. In theory, lots of different tools could all rely on the config tree, leveraging the investment made into creating and updating a config tree. In practice, a fair amount of code that belongs in the parser ended up in the Cricket application itself, which means that it will take some reengineering to make the dream of the One True Config Tree come true.

The User Interface

The fastest, most flexible data collection on the planet doesn't help at all if there's no way to see the data. MRTG led us down the right path here; the graphs it makes are simple and useful, and they are presented in a web page viewable on virtually any platform. For Cricket, we decided to make the HTML and the graphs on the fly, instead of caching completed graphs like MRTG does. This saved some of the processing overhead inherent in MRTG's regular polling cycle. As a bonus, using the CGI script instead of pre-generated HTML soundly defeats the over-aggressive browser-side caching that plagues many MRTG installations. On the down side, the startup time for the CGI script can be a problem, as can the graph generation time, both of which the user perceives as lag in the user interface. The startup time problems could be mitigated by the mod_perl Apache module, and Cricket already uses a graph caching scheme to attempt to avoid wasting time re-rendering graphs. It's generally true, though, that the Cricket user interface is slower than the MRTG interface.

The CGI graph browser was specifically designed to maximize flexibility. It is read-only, which keeps security concerns to a minimum. If access control to sensitive information is required, it is currently only possible on a per-installation basis. This is one drawback to the CGI graph browser, but it has not been a problem for us. There are also ways to run a graph browser on a subset of all the collected data, so access controls could probably be implemented by a determined Cricket user. The graph browser operates entirely on URL's which can be incorporated into web pages outside the Cricket system, or stored as bookmarks. This has made it possible to build ``views'' of the Cricket data which are annotated in various ways to assist operations personnel. We can also use these external pages to collect a set of interesting graphs which might be difficult to navigate among inside the graph browser.

To aid in analyzing the data, it can be useful to assemble graphs out of various data streams which are separately collected and stored by Cricket. So far, this capability is limited to putting multiple graphs on one page, and gathering data from several sources and plotting it in summary on a single graph. These advanced features intended to make it easier to look at the data are a tough sell; they tend to be difficult to implement within the current design, and are only used when trying to solve a particular kind of problem. Of course, having the right tool for the job is invaluable, so it's tempting to try to make Cricket do everything under the sun.

While Cricket is the right tool for these analysis jobs, it's not yet infinitely flexible. It presents useful data that answers 90% of the questions, but the jury's still out on the costs and benefits of adding the complexity necessary to answer the last 9% of the questions. What about the last 1%? We'll never get there, and we don't plan to try. Sometimes, the data suggests questions that are better answered with other tools. The trick is knowing when to turn elsewhere to get your answer.

Instance Mapping

MRTG relies on SNMP instance numbers to identify which of the interfaces on a given router it will pick up data from. One of the things we knew we hated about MRTG was the necessity of re-adjusting instance numbers after any router configuration change. It was a small, but persistent, pain in the neck. It's a dreadfully mundane job, and especially annoying on the large routers we use, which have lots of interfaces. Still, it didn't happen that often, and it hardly showed up on the radar screen of ``things we want to improve in MRTG''

To be fair, the instance number problem is not even really MRTG's fault. It's due to an annoying compromise in the SNMP specification that lets network devices get away with murder in the interest of making them easier to implement. This was done under the assumption that polling systems will be smarter than network devices, and can make up for the lack of intelligence in the managed nodes. SNMP devices are allowed to renumber their interfaces behind the poller's back potentially between every single polling cycle. What's a poller to do? The system can either force humans to fix things up when necessary, or it can attempt to figure out the right instance number on it's own. MRTG uses the former strategy, while Cricket has a feature called ``instance mapping'' to implement the latter strategy.

Cricket's instance mapping feature allows administrators to configure a target by name, and then let Cricket do the necessary SNMP lookups to map that name into an instance number. From then on, it's Cricket's job to keep the instance number correct, no matter how hard the device tries to confuse it. Cricket does this magic in a bulletproof but efficient way. To begin with, it uses a series of SNMP GET-NEXT operations (i.e. a table walk) to map the instance name to a number. Once it has a valid instance number, it caches it. The next time it uses the cached instance number, it fetches the name again, along with the polled data. If a ``no such instance'' error is returned, or the name no longer matches, Cricket discards the polled data, maps the name to a new instance number, and fetches the data again. The new instance number is cached, and will be used until the next re-mapping is required. Fetching the name every time like this amounts to a small overhead in the polling transaction, but it ensures that re-mapping happens as soon as necessary, without any human interaction. Presumably commercial systems solve this problem the same way.

It turns out that this is not just a nice feature, it's a required one in some cases. It was designed to be extremely flexible, since I hope to use Cricket to monitor items in the Host MIB which tend to get different instances all the time (disks and processes, among others). To this end, Cricket can even match against names which are regular expressions, making it easy to find a given process, no matter how it was invoked. I have also been told that the virtual interfaces created and destroyed by Cisco's ATM LANE implementation come and go with unpredictable names, making them very hard to monitor with conventional tools. A Cricket user managed to monitor these virtual interfaces for the first time ever using the instance mapping code and regular expression matching. His exact comment on the triumph was ``Instance mapping is the best thing since sliced bread!''

Future Directions

Our hope for Cricket is that it finds a place in other operations centers, and that it helps improve customer service all around. Poor service reflects poorly on all of us, and it's in the industry's best interests to develop useful tools to raise Quality of Service.

The major features we needed from Cricket are in place now, and it has superseded MRTG at WebTV Networks. The bulk of the work is done. However, there is the usual laundry list of ``todo'' items for Cricket. We will do ongoing work on it, especially to integrate patches from interested helpers.

As we have come to depend on graphs to tell us about the state of the WebTV Service, we have stumbled across a principle that should have been obvious to start with. Simply put, lots of graphs are not a proactive tool. It's not reasonable to expect humans to continually review each of the over 4000 graphs we can produce and watch for anomalies. Of course, Cricket was never intended to simply find links that are down, or hosts that are crashed. There are other tools that we already use for that kind of monitoring. Subtle problems in the system manifest themselves as subtle changes in the patterns on the graph. It is these problems that we seek to discover when we humans browse the graphs. We are looking for graphs that don't ``look right''

The obvious solution to the problem of too many graphs is to ask Cricket itself to evaluate the graphs and flag those that don't ``look right'' for further analysis by a human. This is about the time things get really complicated; we could apply all the fancy pattern recognition tools that the computer science community has to offer, including neural networks, signal processing, and fuzzy logic. We took a more pragmatic view and asked the question, ``How do we humans know when it looks right?'' The answer in all cases was that we compare the current behavior to past behavior. Now the problem looks a bit more solvable. We want to create a system of thresholds which are set differently at different times of day. These thresholds will be derived from past data, so that deviation from past patterns results in violated thresholds, which in turn results in an alert that a human sees. With an appropriate GUI, administrators could adjust the suggested thresholds to avoid false positives. They could also mark known anomalous data with the GUI, so that future generated thresholds would not be affected by the ``bad'' data.

We have asked a group of students from Harvey Mudd College to work on this problem as part of the Computer Science Clinic program. The Clinic program offers a team of undergraduates a chance to work on a real-world problem, while offering companies like WebTV Networks access to bright students who can pursue a project independently. I work less than an hour a week with the team as a liaison to help them understand the problem and our expectations. In return, they are responsible for managing themselves and delivering a finished product at the end of the school year.

The threshold system the clinic team is working on will store its data in the config tree. It will take advantage of inheritance to eliminate duplicated thresholds. It will have a GUI that can be used to look at and modify the suggested time-based threshold curves which are generated using historical data. Threshold violations that are discovered by the collector as it fetches the data will either spawn a shell script, or forward the alert to an alert management system via SNMP traps. The result of the project will be incorporated into the standard Cricket distribution later this year. It will be available under the same license as Cricket itself, the GNU Public License (GPL).

Of course, the real future for Cricket lies with all of the other folks who pick it up and use it. Because it is distributed in source form, and is protected by the GPL, Cricket users are guaranteed the right to hack on it however they see fit. If you need Cricket to do something it cannot already do, you can write the code and share it as contributed software or as a patch. I'm looking forward to hearing from Cricket users about how it's working for them, and seeing what features they need to add to make it work even better.


The Cricket homepage is at: http://cricket.sf.net/.

There's more information there, including the Cricket distribution, where to get the other things you need to make Cricket work, and what mailing lists you might want to join for help. People who want to get in touch with me personally can send e-mail to jra@users.sf.net.


Cricket would not have been written if WebTV Networks had never gotten so addicted to MRTG. So hats off to Tobias Oetiker for a great tool, and for coming through with RRD, a perfect encore. Lots of the ideas for how to improve Cricket came from Jeff Jensen, one of the network administrators at WebTV Networks. The WebTV Networks management lets me spend time on this project, and chose to make it freely available on the net, to boot. Laura de Leon reviewed this paper and helped me spot some unanswered questions. Thanks to you all for your help.

End Notes

1. Actually, this was before I arrived at WebTV Networks. Joe Balboa originally brought MRTG to our company.

2. More information about MRTG is available from http://www.mrtg.org.

3. Writing code is another issue entirely. As a toolsmith, my job is to write the tools WebTV employees need to do their work. Much as they might like to, the network administrators at WebTV Networks don't normally have the time to hack the Cricket code.

4. The RRD website is at: http://www.rrdtool.org/.