Fastest way to search in a string collection

Problem:

I have a text file of around 120,000 users (strings) which I would like to store in a collection and later to perform a search on that collection.

The search method will occur every time the user change the text of a TextBox and the result should be the strings that contain the text in TextBox.

I don't have to change the list, just pull the results and put them in a ListBox.

What I've tried so far:

I tried with two different collections/containers, which I'm dumping the string entries from an external text file (once, of course):

  1. List<string> allUsers;
  2. HashSet<string> allUsers;

With the following LINQ query:

allUsers.Where(item => item.Contains(textBox_search.Text)).ToList();

My search event (fires when user change the search text):

private void textBox_search_TextChanged(object sender, EventArgs e)
{
if (textBox_search.Text.Length > 2)
{
listBox_choices.DataSource = allUsers.Where(item => item.Contains(textBox_search.Text)).ToList();
}
else
{
listBox_choices.DataSource = null;
}
}

Results:

Both gave me a poor response time (around 1-3 seconds between each key press).

Question:

Where do you think my bottleneck is? The collection I've used? The search method? Both?

How can I get better performance and more fluent functionality?

26190 次浏览

Run the search on another thread, and show some loading animation or a progress bar while that thread is running.

You may also try to parallelize the LINQ query.

var queryResults = strings.AsParallel().Where(item => item.Contains("1")).ToList();

Here is a benchmark that demonstrates the performance advantages of AsParallel():

{
IEnumerable<string> queryResults;
bool useParallel = true;


var strings = new List<string>();


for (int i = 0; i < 2500000; i++)
strings.Add(i.ToString());


var stp = new Stopwatch();


stp.Start();


if (useParallel)
queryResults = strings.AsParallel().Where(item => item.Contains("1")).ToList();
else
queryResults = strings.Where(item => item.Contains("1")).ToList();


stp.Stop();


Console.WriteLine("useParallel: {0}\r\nTime Elapsed: {1}", useParallel, stp.ElapsedMilliseconds);
}

It might also be useful to have a "debounce" type of event. This differs from throttling in that it waits a period of time (for example, 200 ms) for changes to finish before firing the event.

See Debounce and Throttle: a visual explanation for more information about debouncing. I appreciate that this article is JavaScript focused, instead of C#, but the principle applies.

The advantage of this is that it doesn't search when you're still entering your query. It should then stop trying to perform two searches at once.

You could try using PLINQ (Parallel LINQ). Although this does not garantee a speed boost, this you need to find out by trial and error.

I doubt you'll be able to make it faster, but for sure you should:

a) Use the AsParallel LINQ extension method

a) Use some kind of timer to delay filtering

b) Put a filtering method on another thread

Keep some kind of string previousTextBoxValue somewhere. Make a timer with a delay of 1000 ms, that fires searching on tick if previousTextBoxValue is same as your textbox.Text value. If not - reassign previousTextBoxValue to the current value and reset the timer. Set the timer start to the textbox changed event, and it'll make your application smoother. Filtering 120,000 records in 1-3 seconds is OK, but your UI must remain responsive.

First I would change how ListControl sees your data source, you're converting result IEnumerable<string> to List<string>. Especially when you just typed few characters this may be inefficient (and unneeded). Do not make expansive copies of your data.

  • I would wrap .Where() result to a collection that implements only what is required from IList (search). This will save you to create a new big list for each character is typed.
  • As alternative I would avoid LINQ and I'd write something more specific (and optimized). Keep your list in memory and build an array of matched indices, reuse array so you do not have to reallocate it for each search.

Second step is to do not search in the big list when small one is enough. When user started to type "ab" and he adds "c" then you do not need to research in the big list, search in the filtered list is enough (and faster). Refine search every time is possible, do not perform a full search each time.

Third step may be harder: keep data organized to be quickly searched. Now you have to change the structure you use to store your data. imagine a tree like this:

A        B         C
Add      Better    Ceil
Above    Bone      Contour

This may simply be implemented with an array (if you're working with ANSI names otherwise a dictionary would be better). Build the list like this (illustration purposes, it matches beginning of string):

var dictionary = new Dictionary<char, List<string>>();
foreach (var user in users)
{
char letter = user[0];
if (dictionary.Contains(letter))
dictionary[letter].Add(user);
else
{
var newList = new List<string>();
newList.Add(user);
dictionary.Add(letter, newList);
}
}

Search will be then done using first character:

char letter = textBox_search.Text[0];
if (dictionary.Contains(letter))
{
listBox_choices.DataSource =
new MyListWrapper(dictionary[letter].Where(x => x.Contains(textBox_search.Text)));
}

Please note I used MyListWrapper() as suggested in first step (but I omitted by 2nd suggestion for brevity, if you choose right size for dictionary key you may keep each list short and fast to - maybe - avoid anything else). Moreover note that you may try to use first two characters for your dictionary (more lists and shorter). If you extend this you'll have a tree (but I don't think you have such big number of items).

There are many different algorithms for string searching (with related data structures), just to mention few:

  • Finite state automaton based search: in this approach, we avoid backtracking by constructing a deterministic finite automaton (DFA) that recognizes stored search string. These are expensive to construct—they are usually created using the powerset construction—but are very quick to use.
  • Stubs: Knuth–Morris–Pratt computes a DFA that recognizes inputs with the string to search for as a suffix, Boyer–Moore starts searching from the end of the needle, so it can usually jump ahead a whole needle-length at each step. Baeza–Yates keeps track of whether the previous j characters were a prefix of the search string, and is therefore adaptable to fuzzy string searching. The bitap algorithm is an application of Baeza–Yates' approach.
  • Index methods: faster search algorithms are based on preprocessing of the text. After building a substring index, for example a suffix tree or suffix array, the occurrences of a pattern can be found quickly.
  • Other variants: some search methods, for instance trigram search, are intended to find a "closeness" score between the search string and the text rather than a "match/non-match". These are sometimes called "fuzzy" searches.

Few words about parallel search. It's possible but it's seldom trivial because overhead to make it parallel can be easily much higher that search itself. I wouldn't perform search itself in parallel (partitioning and synchronization will become soon too expansive and maybe complex) but I would move search to a separate thread. If main thread isn't busy your users won't feel any delay while they're typing (they won't note if list will appear after 200 ms but they'll feel uncomfortable if they have to wait 50 ms after they typed). Of course search itself must be fast enough, in this case you don't use threads to speed up search but to keep your UI responsive. Please note that a separate thread will not make your query faster, it won't hang UI but if your query was slow it'll still be slow in a separate thread (moreover you have to handle multiple sequential requests too).

Assuming you are only matching by prefixes, the data structure you are looking for is called a trie, also known as "prefix tree". The IEnumerable.Where method that you're using now will have to iterate through all items in your dictionary on each access.

This thread shows how to create a trie in C#.

You need either a text search engine (like Lucene.Net), or database (you may consider an embedded one like SQL CE, SQLite, etc.). In other words, you need an indexed search. Hash-based search isn't applicable here, because you searching for sub-string, while hash-based search is well for searching for exact value.

Otherwise it will be an iterative search with looping through the collection.

I've done some testing, and searching a list of 120,000 items and populating a new list with the entries takes a negligible amount of time (about a 1/50th of a second even if all strings are matched).

The problem you're seeing must therefore be coming from the populating of the data source, here:

listBox_choices.DataSource = ...

I suspect you are simply putting too many items into the listbox.

Perhaps you should try limiting it to the first 20 entries, like so:

listBox_choices.DataSource = allUsers.Where(item => item.Contains(textBox_search.Text))
.Take(20).ToList();

Also note (as others have pointed out) that you are accessing the TextBox.Text property for each item in allUsers. This can easily be fixed as follows:

string target = textBox_search.Text;
listBox_choices.DataSource = allUsers.Where(item => item.Contains(target))
.Take(20).ToList();

However, I timed how long it takes to access TextBox.Text 500,000 times and it only took 0.7 seconds, far less than the 1 - 3 seconds mentioned in the OP. Still, this is a worthwhile optimisation.

You could consider doing the filtering task on a background thread which would invoke a callback method when it's done, or simply restart filtering if input is changed.

The general idea is to be able to use it like this:

public partial class YourForm : Form
{
private readonly BackgroundWordFilter _filter;


public YourForm()
{
InitializeComponent();


// setup the background worker to return no more than 10 items,
// and to set ListBox.DataSource when results are ready


_filter = new BackgroundWordFilter
(
items: GetDictionaryItems(),
maxItemsToMatch: 10,
callback: results =>
this.Invoke(new Action(() => listBox_choices.DataSource = results))
);
}


private void textBox_search_TextChanged(object sender, EventArgs e)
{
// this will update the background worker's "current entry"
_filter.SetCurrentEntry(textBox_search.Text);
}
}

A rough sketch would be something like:

public class BackgroundWordFilter : IDisposable
{
private readonly List<string> _items;
private readonly AutoResetEvent _signal = new AutoResetEvent(false);
private readonly Thread _workerThread;
private readonly int _maxItemsToMatch;
private readonly Action<List<string>> _callback;


private volatile bool _shouldRun = true;
private volatile string _currentEntry = null;


public BackgroundWordFilter(
List<string> items,
int maxItemsToMatch,
Action<List<string>> callback)
{
_items = items;
_callback = callback;
_maxItemsToMatch = maxItemsToMatch;


// start the long-lived backgroud thread
_workerThread = new Thread(WorkerLoop)
{
IsBackground = true,
Priority = ThreadPriority.BelowNormal
};


_workerThread.Start();
}


public void SetCurrentEntry(string currentEntry)
{
// set the current entry and signal the worker thread
_currentEntry = currentEntry;
_signal.Set();
}


void WorkerLoop()
{
while (_shouldRun)
{
// wait here until there is a new entry
_signal.WaitOne();
if (!_shouldRun)
return;


var entry = _currentEntry;
var results = new List<string>();


// if there is nothing to process,
// return an empty list
if (string.IsNullOrEmpty(entry))
{
_callback(results);
continue;
}


// do the search in a for-loop to
// allow early termination when current entry
// is changed on a different thread
foreach (var i in _items)
{
// if matched, add to the list of results
if (i.Contains(entry))
results.Add(i);


// check if the current entry was updated in the meantime,
// or we found enough items
if (entry != _currentEntry || results.Count >= _maxItemsToMatch)
break;
}


if (entry == _currentEntry)
_callback(results);
}
}


public void Dispose()
{
// we are using AutoResetEvent and a background thread
// and therefore must dispose it explicitly
Dispose(true);
}


private void Dispose(bool disposing)
{
if (!disposing)
return;


// shutdown the thread
if (_workerThread.IsAlive)
{
_shouldRun = false;
_currentEntry = null;
_signal.Set();
_workerThread.Join();
}


// if targetting .NET 3.5 or older, we have to
// use the explicit IDisposable implementation
(_signal as IDisposable).Dispose();
}
}

Also, you should actually dispose the _filter instance when the parent Form is disposed. This means you should open and edit your Form's Dispose method (inside the YourForm.Designer.cs file) to look something like:

// inside "xxxxxx.Designer.cs"
protected override void Dispose(bool disposing)
{
if (disposing)
{
if (_filter != null)
_filter.Dispose();


// this part is added by Visual Studio designer
if (components != null)
components.Dispose();
}


base.Dispose(disposing);
}

On my machine, it works pretty quickly, so you should test and profile this before going for a more complex solution.

That being said, a "more complex solution" would possibly be to store the last couple of results in a dictionary, and then only filter them if it turns out that the new entry differs by only the first of last character.

Update:

I did some profiling.

(Update 3)

  • List content: Numbers generated from 0 to 2.499.999
  • Filter text: 123 (20.477 results)
  • Core i5-2500, Win7 64bit, 8GB RAM
  • VS2012 + JetBrains dotTrace

The initial test run for 2.500.000 records took me 20.000ms.

Number one culprit is the call to textBox_search.Text inside Contains. This makes a call for each element to the expensive get_WindowText method of the textbox. Simply changing the code to:

    var text = textBox_search.Text;
listBox_choices.DataSource = allUsers.Where(item => item.Contains(text)).ToList();

reduced the execution time to 1.858ms.

Update 2 :

The other two significant bottle-necks are now the call to string.Contains (about 45% of the execution time) and the update of the listbox elements in set_Datasource (30%).

We could make a trade-off between speed and memory usage by creating a Suffix tree as Basilevs has suggested to reduce the number of necessary compares and push some processing time from the search after a key-press to the loading of the names from file which might be preferable for the user.

To increase the performance of loading the elements into the listbox I would suggest to load only the first few elements and indicate to the user that there are further elements available. This way you give a feedback to the user that there are results available so they can refine their search by entering more letters or load the complete list with a press of a button.

Using BeginUpdate and EndUpdate made no change in the execution time of set_Datasource.

As others have noted here, the LINQ query itself runs quite fast. I believe your bottle-neck is the updating of the listbox itself. You could try something like:

if (textBox_search.Text.Length > 2)
{
listBox_choices.BeginUpdate();
listBox_choices.DataSource = allUsers.Where(item => item.Contains(textBox_search.Text)).ToList();
listBox_choices.EndUpdate();
}

I hope this helps.

Use Suffix tree as index. Or rather just build a sorted dictionary that associates every suffix of every name with the list of corresponding names.

For input:

Abraham
Barbara
Abram

The structure would look like:

a -> Barbara
ab -> Abram
abraham -> Abraham
abram -> Abram
am -> Abraham, Abram
aham -> Abraham
ara -> Barbara
arbara -> Barbara
bara -> Barbara
barbara -> Barbara
bram -> Abram
braham -> Abraham
ham -> Abraham
m -> Abraham, Abram
raham -> Abraham
ram -> Abram
rbara -> Barbara

Search algorithm

Assume user input "bra".

  1. Bisect the dictionary on user input to find the user input or the position where it could go. This way we find "barbara" - last key lower than "bra". It is called lower bound for "bra". Search will take logarithmic time.
  2. Iterate from the found key onwards until user input no longer matches. This would give "bram" -> Abram and "braham" -> Abraham.
  3. Concatenate iteration result (Abram, Abraham) and output it.

Such trees are designed for quick search of substrings. It performance is close to O(log n). I believe this approach will work fast enough to be used by GUI thread directly. Moreover it will work faster then threaded solution due to absence of synchronization overhead.

I would try to sort collection, search to match only start part and limit search by some number.

so on ininialization

allUsers.Sort();

and search

allUsers.Where(item => item.StartWith(textBox_search.Text))

Maybe you can add some cache.

Use Parallel LINQ. PLINQ is a parallel implementation of LINQ to Objects. PLINQ implements the full set of LINQ standard query operators as extension methods for the T:System.Linq namespace and has additional operators for parallel operations. PLINQ combines the simplicity and readability of LINQ syntax with the power of parallel programming. Just like code that targets the Task Parallel Library, PLINQ queries scale in the degree of concurrency based on the capabilities of the host computer.

Introduction to PLINQ

Understanding Speedup in PLINQ

Also you can use Lucene.Net

Lucene.Net is a port of the Lucene search engine library, written in C# and targeted at .NET runtime users. The Lucene search library is based on an inverted index. Lucene.Net has three primary goals:

The WinForms ListBox control really is your enemy here. It will be slow to load the records and the ScrollBar will fight you to show all 120,000 records.

Try using an old-fashioned DataGridView data-sourced to a DataTable with a single column [UserName] to hold your data:

private DataTable dt;


public Form1() {
InitializeComponent();


dt = new DataTable();
dt.Columns.Add("UserName");
for (int i = 0; i < 120000; ++i){
DataRow dr = dt.NewRow();
dr[0] = "user" + i.ToString();
dt.Rows.Add(dr);
}
dgv.AutoSizeColumnsMode = DataGridViewAutoSizeColumnsMode.Fill;
dgv.AllowUserToAddRows = false;
dgv.AllowUserToDeleteRows = false;
dgv.RowHeadersVisible = false;
dgv.DataSource = dt;
}

Then use a DataView in the TextChanged event of your TextBox to filter the data:

private void textBox1_TextChanged(object sender, EventArgs e) {
DataView dv = new DataView(dt);
dv.RowFilter = string.Format("[UserName] LIKE '%{0}%'", textBox1.Text);
dgv.DataSource = dv;
}

According to what I have seen I agree with the fact to sort the list.

However to sort when the list is construct will be very slow, sort when building, you will have a better execution time.

Otherwise if you don't need to display the list or to keep the order, use a hashmap.

The hashmap will hash your string and search at the exact offset. It should be faster I think.

You can also try using BindingSource.Filter function. I have used it and it works like a charm to filter from bunch of records, every time update this property with the text being search. Another option would be to use AutoCompleteSource for TextBox control.

Hope it helps!

Try use BinarySearch method it should work faster then Contains method.

Contains will be an O(n) BinarySearch is an O(lg(n))

I think that sorted collection should work faster on search and slower on adding new elements, but as I understood you have only search perfomance problem.