Updated with new reference material May 4th 2009
How NetEqualizer compares to Packeteer, Allot, Cymphonics, Exinda
We often get asked how NetEqualizer compares to Packeteer, Allot, Cymphonics, Exinda and a plethora of other well-known companies that do layer 7 application shaping (packet shaping). After several years of these questions, and discussing different aspects with former and current application shaping IT administrators, we’ve developed a response that should clarify the differences between NetEqualizers behavior based approach and the rest of the pack.
We thought of putting our response into a short, bullet-by-bullet table format, but then decided that since this decision often involves tens of thousands of dollars, 15 minutes of education on the subject with content to support the bullet chart was in order. If you want to see just the bullet chart, you can skip to the end now, but if you’re looking to have the question answered as objectively as possible, please take a few minutes to read on
In the following sections, we will cover specifically when and where application shaping (deep packet inspection) is used, how it can be used to your advantage, and also when it may not be a good option for what you are trying to accomplish. We will also discuss how the NetEqualizer and its behavior-based shaping fits into the landscape of application shaping, and how in some cases the NetEqualizer is a much better alternative.
First off, let’s discuss the accuracy of application shaping. To do this, we need to review the basic mechanics of how it works.
Application shaping is defined as the ability to identify traffic on your network by type and then set customized policies to control the flow rates for each particular type. For example, Citrix, AIM, Youtube, and BearShare are all applications that can be uniquely identified.
As you are likely aware, all traffic on the Internet travels around in what is called an IP packet. An IP packet can very simply be thought of as a string of characters moving from computer A to computer B. The string of characters is called the “payload,” much like the freight inside a railroad car. On the outside of this payload is the address where it is being sent. On the inside is the data/payload that is being transmitted. These two elements, the address and the payload, comprise the complete IP packet. In the case of different applications on the Internet, we would expect to see different kinds of payloads.
At the heart of all current application shaping products is special software that examines the content of Internet packets as they pass through the packet shaper. Through various pattern matching techniques, the packet shaper determines in real time what type of application a particular flow is. It then proceeds to take action to possibly restrict or allow the data based on a rule set designed by the system administrator.
For example, the popular peer-to-peer application Kazaa actually has the ASCII characters “Kazaa” appear in the payload, and hence a packet shaper can use this keyword to identify a Kazaa application. Seems simple enough, but suppose that somebody was downloading a Word document discussing the virtues of peer-to-peer and the title had the character string “Kazaa” in it. Well, it is very likely that this download would be identified as Kazaa and hence misclassified. After all, downloading a Word document from a Web server is not the same thing as the file sharing application Kazaa.
The other issue that constantly brings the accuracy of application shaping under fire is that some application writers find it in their best interest not be classified. In a mini arms race that plays out everyday across the world, some application developers are constantly changing their signature and some have gone as far as to encrypt their data entirely.
Yes, it is possible for the makers of application shapers to counter each move, and that is exactly what the top companies do, but it can take a heroic effort to keep pace. The constant engineering and upgrading required has an escalating cost factor. In the case of encrypted applications, the amount of CPU power required for decryption is quite intensive and impractical and other methods will be needed to identify encrypted p2p.
But, this is not to say that application shaping doesn’t work in all cases or provide some value. So, let’s break down where it has potential and where it may bring false promises. First off, the realities of what really happens when you deploy and depend on this technology need to be discussed.
Accuracy and False Positives
As of early 2003, we had a top engineer and executive join APConnections direct from a company that offered application shaping as one of their many value-added technologies. He had first hand knowledge from working with hundreds of customers who were big supporters of application shaping:
The application shaper his company offered could identify 90 percent of the spectrum of applications, which means they left 10 percent as unclassified. So, right off the bat, 10 percent of the traffic is unknown by the traffic shaper. Is this traffic important? Is it garbage that you can ignore? Well, there is no way to know with out any intelligence, so you are forced to let it go by without any restriction. Or, you could put one general rule over all of the traffic – perhaps limiting it to 1 megabit per second max, for example. Essentially, if your intention was 100-percent understanding and control of your network traffic, right out the gate you must compromise this standard.
In fairness, this 90-percent identification actually is an amazing number with regard to accuracy when you understand how daunting application shaping is. Regardless, there is still room for improvement.
So, that covers the admitted problem of unclassifiable traffic, but how accurate can a packet shaper be with the traffic it does claim to classify? Does it make mistakes? There really isn’t any reliable data on how often an application shaper will misidentify an application. To our knowledge, there is no independent consumer reporting company that has ever created a lab capable of generating several thousand different applications types with a mix of random traffic, and then took this mix and identified how often traffic was misclassified. Yes, there are trivial tests done one application at a time, but misclassification becomes more likely with real-world complex and diverse application mixes.
From our own testing of application technology freely available on the Internet, we discovered false positives can occur up to 25 percent of the time. A random FTP file download can be classified as something more specific. Obviously commercial packet shapers do not rely on the free technology in open source and they actually may improve on it. So, if we had to estimate based on our experience, perhaps 5 percent of Internet traffic will likely get misclassified. This brings our overall accuracy down to 85 percent (combining the traffic they don’t claim to classify with an estimated error rate for the traffic they do classify).
Constantly Evolving Traffic
Our sources say (mentioned above) that 70 percent of their customers that purchased application shaping equipment were using the equipment primarily as a reporting tool after one year. This means that they had stopped keeping up with shaping policies altogether and were just looking at the reports to understand their network (nothing proactive to change the traffic).
This is an interesting fact. From what we have seen, many people are just unable, or unwilling, to put in the time necessary to continuously update and change their application rules to keep up with the evolving traffic. The reason for the constant changing of rules is that with traditional application shaping you are dealing with a cunning and wise foe. For example, if you notice that there is a large contingent of users using Bittorrent and you put a rule in to quash that traffic, within perhaps days, those users will have moved on to something new: perhaps a new application or encrypted p2p. If you do not go back and reanalyze and reprogram your rule set, your packet shaper slowly becomes ineffective.
And finally lest we not forget that application shaping is considered by some to be a a violation of Net Neutrality.
When is application shaping the right solution?
There is a large set of businesses that use application shaping quite successfully along with other technologies. This area is WAN optimization. Thus far, we have discussed the issues with using an application shaper on the wide open Internet where the types and variations of traffic are unbounded. However, in a corporate environment with a finite set and type of traffic between offices, an application shaper can be set up and used with fantastic results.
There is also the political side to application shaping. It is human nature to want to see and control what takes place in your environment. Finding the best tool available to actually show what is on your network, and the ability to contain it, plays well with just about any CIO or IT director on the planet. An industry leading packet shaper brings visibility to your network and a pie chart showing 300 different kinds of traffic. Whether or not the tool is practical or accurate over time isn’t often brought into the buying decision. The decision to buy can usually be “intuitively” justified. By intuitively, we mean that it is easier to get approval for a tool that is simple to conceptually understand by a busy executive looking for a quick-fix solution.
As the cost of bandwidth continues to fall, the question becomes how much a CIO should spend to analyze a network. This is especially true when you consider that as the Internet expands, the complexity of shaping applications grows. As bandwidth prices drop, the cost of implementing such a product is either flat or increasing. In cases such as this, it often does not make sense to purchase a $15,000 bandwidth shaper to stave off a bandwidth upgrade that might cost an additional $200 a month.
What about the reporting aspects of an application shaper? Even if it can only accurately report 90 percent of the actual traffic, isn’t this useful data in itself?
Yes and no. Obviously analyzing 90 percent of the data on your network might be useful, but if you really look at what is going on, it is hard to feel like you have control or understanding of something that is so dynamic and changing. By the time you get a handle on what is happening, the system has likely changed. Unless you can take action in real time, the network usage trends (on a wide open Internet trunk) will vary from day to day.1 It turns out that the most useful information you can determine regarding your network is an overall usage patter for each individual. The goof-off employee/user will stick out like a sore thumb when you look at a simple usage report since the amount of data transferred can be 10-times the average for everybody else. The behavior is the indicator here, but the specific data types and applications will change from day to day and week to week
How does the NetEqualizer differ and what are its advantages and weaknesses?
First, we’ll summarize equalizing and behavior-based shaping. Overall, it is a simple concept. Equalizing is the art form of looking at the usage patterns on the network, and then when things get congested, robbing from the rich to give to the poor. Rather than writing hundreds of rules to specify allocations to specific traffic as in traditional application shaping, you can simply assume that large downloads are bad, short quick traffic is good, and be done with it.
This behavior-based approach usually mirrors what you would end up doing if you could see and identify all of the traffic on your network, but doesn’t require the labor and cost of classifying everything. Applications such as Web surfing, IM, short downloads, and voice all naturally receive higher priority while large downloads and p2p receive lower priority. This behavior-based shaping does not need to be updated constantly as applications change.
Trusting a heuristic solution such as NetEqualizer is not always an easy step. Oftentimes, customers are concerned with accidentally throttling important traffic that might not fit the NetEqualizer model, such as video. Although there are exceptions, it is rare for the network operator not to know about these potential issues in advance, and there are generally relatively few to consider. In fact, the only exception that we run into is video, and the NetEqualizer has a low level routine that easily allows you to give overriding priority to a specific server on your network, hence solving the problem.
Another key element in behavior-based shaping is connections. Equalizing takes care of instances of congestion caused by single-source bandwidth hogs. However, the other main cause of Internet gridlock (as well as bringing down routers and access points) is p2p and its propensity to open hundreds or perhaps thousands of connections to different sources on the Internet. Over the years, the NetEqualizer engineers have developed very specific algorithms to spot connection abuse and avert its side effects.
This overview, along with the summary table below, should give you a good idea of where the NetEqualizer stands in relation to packet shaping.
Application based shaping
Deep Packet Inspection Abuse In Iran Raises Questions About DPI WorldwideJuly 10, 2009 — netequalizer
Over the past few years, we at APconnections have made our feelings about Deep Packet Inspection clear, completely abandoning the practice in our NetEqualizer technology more than two years ago. While there may be times that DPI is necessary and appropriate, its use in many cases can threaten user privacy and the open nature of the Internet. And, in extreme cases, DPI can even be used to threaten freedom of speech and expression. As we mentioned in a previous article, this is currently taking place in Iran.
Although these extreme invasions of privacy are most likely not occurring in the United States, their existence in Iran is bringing increasing attention to the slippery slope that is Deep Packet Inspection. A July 10 Huffington Post article reads:
This potential harm to the privacy and rights of consumers was a major factor behind our decision to discontinue the use of DPI in any of our technology and invest in alternative means for network optimization. We hope that the ongoing controversy will be reason for others to do the same.