Click here to Skip to main content
15,867,756 members
Articles / Programming Languages / C#
Article

Extreme Optimization #1.1: Mapping IP addresses to country codes.

Rate me:
Please Sign up or sign in to vote.
4.81/5 (34 votes)
30 May 200310 min read 121.5K   2.9K   70   14
Highly optimized classes for looking up the country code corresponding to an IP address

Introduction

Extreme Optimization can lead to remarkable improvements in both speed and size. In this article, we will see an example of how we can modify a standard algorithm using our knowledge of the problem at hand to achieve significant performance gains.

Background

We build on an article written by R. Reyes called Optimized IP to ISO3166 Country Code Mapping in C#. The objective is to perform fast lookup of the country code corresponding to an IP address in a table of over 52,000 entries while keeping memory usage to a minimum. We are only interested in the current IPv4 addresses, consisting of a sequence of 4 byte values in the range 0-255.

The existing solution

Mr. Reyes introduces the PATRICIA trie structure (pronounced ‘try’ according to Donald Knuth) to efficiently perform the lookups. We will not repeat the details of the implementation here. They can be found in the original article.

The Extreme Optimization

Extreme Optimization™ begins with a thorough understanding of the problem we are solving. We combine this understanding with in-depth knowledge of existing algorithms

Storing IP addresses more efficiently

The original algorithm uses a BitVector class to store the keys. Keys are represented internally as a Byte array. Only the distinctive portion of the key is stored, so many vectors will actually be only one byte long. However, a lot of shifting of bits is involved in isolating the relevant portion of the key and to check for matches.

A first observation is that we do not need a general BitVector class to store our data. Our keys are sequences of up to 32 bits. They fit perfectly into a 32 bit unsigned integer. This will eliminate the overhead of managing the array object.

Furthermore, we can eliminate the excessive shifting by storing the entire key in every node and using the bitwise Xor operator. The boolean Xor (exclusive OR) operator returns true if its operands are different and false if they are the same. The bitwise version compares identically positioned bits in two numbers. It assigns 0 to the corresponding bit in the result when the bits are the same, and 1 when they are different.

To find the shared portion of two keys, we Xor the two values. The key length is the number of leading zeros in the result. We can find this number by comparing the result to powers of two. For a key length of 32 bits, if the result is smaller than 2^n, we know that the first 32-n bits of the two keys are the same.

Here is an example. We want to find the common key length of two IP addresses, IP1= 207.176.130.34 and IP2= 207.176.151.11.

IP Address Binary value Numerical value

IP1 (207.176.130.34)

11001111.10110000.10000010.00100010

411,303,936

IP2 (207.176.151.11)

11001111.10110000.10010111.00001011

411,336,704

IP1 xor IP2

00000000.00000000.00010101.00101001

5,417

The result, 5417, is smaller than 8192 = 2^13. Therefore, the two IP’s have the first 32 - 13 = 19 bits in common.

When comparing a key value to the key values of the children of a node, we can start counting up from the key length of the parent node.

Optimizing Patricia tries for binary keys

Secondly, we are dealing with binary data. The acronym PATRICIA is short for Practical Algorithm To Retrieve Information Coded In Alphanumeric. This name suggests that Patricia tries are best suited for alphanumerical data, where every node can have many children. The keys in our problem are sequences of 0’s and 1’s. This means that any node can have at most two children. One child will have a prefix starting with a 0. The other will have a prefix starting with a 1.

This means we can eliminate the ArrayList we have been using to store the references to the child nodes. We replace it with two references, one each for the child whose key prefix starts with a 0 or a 1, respectively. If either child is not present, the corresponding reference is null.

An ArrayList is a relatively expensive structure to store very short lists. In addition to the usual overhead associated with reference types, it also allocates space for unused entries. Eliminating both the overhead and the unused space should significantly reduce our memory requirements.

When looking up values in the trie, we can simply use the value of the first bit after the current key length to select the next child node. We can then quickly check that our value matches the full length of the child node.

Further Optimizations

A trie can work with one root element. The first few steps of the lookup will be spent working through the first byte or 8 bits of the IP address. If we create a separate root for each possible value of the first bits in the key, we eliminate 6 or 7 unnecessary recursive steps. We can do even better by pre-selecting on more leading bits. We will call this number of bits the index length.

We have to be somewhat careful, however. Every root node spans a certain range of IP addresses. Some entries in our input tables will span a range that covers more than one root node. We will have to add these entries to all the appropriate root nodes, adjusting the range to fit that root nodes' range. This will come at a small penalty in the memory footprint, but the speed gain is significant.

If we use an index length of at least 1, we can use Int32 values for the keys instead of the non CLS-compliant UInt32 without any complications resulting from using signed numbers. A key passed to a node will always have at least the leading bit in common with the node's key. The leading bit of the result of the XOR operations will therefore always be zero, indicating a positive value.

Looking at the code

The new project contains three classes.

  • BinaryTrie represents a trie with 32 bit binary values as its keys.
  • BinaryTrieNode represents a node in BinaryTrie.
  • IPCountryTable represents a lookup table. It inherits from BinaryTrie.

BinaryTrie has two constructors. The first constructor takes one argument, indexLength, which is the number of bits to use in pre-selection. It creates the _roots array which will contain the root nodes. The default constructor uses an index length of 1.

The Add method adds a key to the trie and returns the BinaryTrieNode corresponding to the key. The AddInternal method does the actual work of adding a key to the trie. If no root node exists for the given index, it creates one. If the root node exists, the request is passed on to the root node, which does the bulk of the work.

BinaryTrieNode represents an entry in the trie. It contains the following fields:

  • _key is a UInt32 value containing the key of the current node.
  • _keyLength is the number of bits in the current key prefix.
  • _userData points to any user-supplied object associated with the key. In this instance, this will be the Country Code associated with the node.
  • _zero contains a reference to the child node whose prefix starts with a 0.
  • _one contains a reference to the child node whose prefix starts with a 1.

BinaryTrieNode has one internal constructor that is used by BinaryTrie to create a root node. It also contains two private constructors. One creates a shallow copy of a node. The other creates a node using the data that was passed to the Add method. It calculates the key length based on the number of addresses available to the network.

Most of the hard work of building the trie is done in the AddInternal() method. We first calculate the common key length of the current node and the new entry. If the new key length is smaller than the common key length, the new entry becomes the parent of the current node. If the match is complete, i.e. the common key length is at least the key length of the current node, the new entry is passed on to a child node. Which child we work with depends on the first bit after the current key.

If the match is incomplete, the current node is replaced with a new node with the common portion of the key as its key. It has two children: (a copy of) the current node, and a node for the new entry. In both cases, the bit at position _keyLength+1 determines which of the two child nodes gets to go in the _zero and _one slots, respectively.

The lookup is passed on from IPCountryTable straight to the correct root BinaryTrieNode, if it exists. If the lookup key matches one of the child node keys, the request is passed on to that child node. Otherwise, we have found the best match.

Benchmark results

Using the same data set and test code as the original article, we found the following results. The Optimized version #1 refers to the original version of the code as posted here on May 13, 2003. The Optimized version #1.1 refers to the latest code, which includes the further optimizations.

Original version

Optimized version #1

Optimized version #1.1

Result

Result

Change

Result

Change

Memory usage

9,409 KB

2,578 KB

- 72.6%

2,963 KB

- 68.5%

Load time

7.361s

2.864s

- 61.1%

1.858s

- 74.8%

25,000 lookups

3.475s

0.2567s

- 92.6%

0.1997s

- 94.3%

Lookups/second

7,194

97,048

+ 1,249%

125,190

+ 1,640%

Here is a graphical representation of the results:

Graphic representation of the benchmark results.

Possible improvements

Is there room for even more optimization? Possibly.

A single country code will have many nodes referencing it. We can eliminate a node if, for each of the IP addresses in its range, the best matching node in the resulting trie has the same country code as the node we eliminated. This will happen if the node's parent node has the same country code.

After the trie has been built, we can go through all the nodes iteratively and eliminate unnecessary nodes. If we can eliminate a significant number of nodes, the memory usage will decrease and the lookup speed will increase accordingly. This possibility will be explored in a future update of this article.

Conclusion

The two main objectives were minimizing memory usage and maximizing speed. We have further optimized an already optimized solution to produce dramatic improvements in both areas. We originally reduced memory usage by 3/4, and increased the lookup speed by a factor of more than 13.

With the latest version, we give up a little bit of memory for an additional 30% increase in performance. We also untangled the binary trie code from the IP lookup code. It can now be used independently.

References

    Update

    APNIC (the Asia Pacific Network Information Centre), one of four organizations that publishes the data used in this article, recently changed the format of its statistics file to conform to the format used by the other three organizations. The size field in the database now contains the total size of the segment rather than the number of bits. To use files in this new format, you will need to make a change in the source code. In the call to LoadStatisticsFile which loads the APNIC data file, set the calculateKeyLength parameter to true.

    License

    This article has no explicit license attached to it but may contain usage terms in the article text or the download files themselves. If in doubt please contact the author via the discussion board below.

    A list of licenses authors might use can be found here


    Written By
    Founder Extreme Optimization
    Canada Canada
    Jeffrey has been writing numerical software for many years. He is founder and president of Extreme Optimization, a Toronto based provider of numerical component libraries for the .NET framework. He loves challenges, especially when it comes to making code run fast, and finding simplicity and elegance in what looks like complicated chaos.

    Comments and Discussions

     
    GeneralWhere could I find the latest source code - With the fix in LoadStatisticsFile() Pin
    aneeshlalga2-Mar-10 23:37
    aneeshlalga2-Mar-10 23:37 
    GeneralTnx for great article Pin
    mikker_12326-Jul-05 5:46
    mikker_12326-Jul-05 5:46 
    GeneralSmall Bug Pin
    Kentamanos12-Jun-03 15:29
    Kentamanos12-Jun-03 15:29 
    GeneralRe: Small Bug Pin
    Jeffrey Sax8-Jul-03 20:45
    Jeffrey Sax8-Jul-03 20:45 
    GeneralUpdate: 30% faster! Pin
    Jeffrey Sax28-May-03 9:43
    Jeffrey Sax28-May-03 9:43 
    GeneralTrie pronounced "Try" not "Tree" Pin
    Blake Coverett27-May-03 17:44
    Blake Coverett27-May-03 17:44 
    GeneralGreat but ... Pin
    Sebastien Lorion25-May-03 19:51
    Sebastien Lorion25-May-03 19:51 
    GeneralRe: Great but ... Pin
    Jeffrey Sax28-May-03 9:45
    Jeffrey Sax28-May-03 9:45 
    GeneralRe: Great but ... Pin
    Sebastien Lorion28-May-03 12:11
    Sebastien Lorion28-May-03 12:11 
    GeneralRe: Great but ... Pin
    Jeffrey Sax28-May-03 12:24
    Jeffrey Sax28-May-03 12:24 
    Questionbugs? Pin
    Frank Hileman22-May-03 4:34
    Frank Hileman22-May-03 4:34 
    AnswerRe: bugs? Pin
    Jeffrey Sax22-May-03 9:47
    Jeffrey Sax22-May-03 9:47 
    GeneralNew apnic.latest has key length pre-calculated Pin
    eekthecat20-May-03 13:35
    eekthecat20-May-03 13:35 
    GeneralRe: New apnic.latest has key length pre-calculated Pin
    Jeffrey Sax20-May-03 14:07
    Jeffrey Sax20-May-03 14:07 

    General General    News News    Suggestion Suggestion    Question Question    Bug Bug    Answer Answer    Joke Joke    Praise Praise    Rant Rant    Admin Admin   

    Use Ctrl+Left/Right to switch messages, Ctrl+Up/Down to switch threads, Ctrl+Shift+Left/Right to switch pages.