This article is the version 2 of my previous article found here (http://www.codeproject.com/Articles/190504/RaptorDB
I had to write a new article because in this version I completely
redesigned and re-architected the original and so it would not go
with the previous article. In this version I have done away with
the b+tree and hash index in favor of my own
which for all intents and purposes is superior and the performance
numbers speak for themselves.
Here is a brief overview of all the
terms used to describe
- Embedded: You can use
your application as you would any other DLL, and you don't
need to install services or run external programs.
- NoSQL: A grass roots movement to replace
relational databases with more relevant and specialized
storage systems to the application in question. These systems
are usually designed for performance.
- Persisted: Any changes made are stored on
hard disk, so you never lose data on power outages or crashes.
- Dictionary: A key/value storage system much
like the implementation in .NET.
- MurMurHash: A non cryptographic hash
function created by Austin Appleby in 2008 (http://en.wikipedia.org/wiki/MurmurHash).
has the following features :
- Very fast performance (typically 2x the insert and 4x the
read performance of
- Extremely small foot print at ~50kb.
- No dependencies.
- Multi-Threaded support for read and writes.
- Data pages are separate from the main tree structure, so can
be freed from memory if needed, and loaded on demand.
- Automatic index file recovery on non-clean shutdowns.
- String Keys are UTF8 encoded and limited to 60 bytes if not
specified otherwise (maximum is 255 chars).
- Support for long string Keys with the
- Duplicate keys are stored as a WAH Bitmap Index for optimal
storage and speed in access.
- Two mode of operation Flush immediate and Deferred ( the
latter being faster at the expense of the risk of non-clean
shutdown data loss).
- Enumerate the index is supported.
- Enumerate the Storage file is supported.
- Remove Key is supported.
Why another data
There is always room for improvement, and the ever need for faster
systems compels us to create new methods of doing things.
is no exception to this rule. Currently
by a factor of 15x on writes
and 21x on reads
, while keeping the
main feature of disk friendliness of a b+tree structure.
The problem with
Theoretically a b+tree is O(N log k N) or log
base k of N, now for the typical values of k which are above 200
for example the b+tree should outperform any binary tree because
it will use less operations. However I have found the following
problems which hinder performance :
- Pages in a b+tree are usually implemented as a list or array
of child pointers and so while finding and inserting a value
is a O(log k) operation the process actually has to move
children around in the array or list, and so is time
- Splitting a page in b+tree has to fix parent nodes and
children so effectively will lock the tree for the duration,
so parallel updates are very very difficult and have spawned a
lot of research articles.
a good index structure
So what makes a good index structure, here are what I consider
essential features of one:
- Page-able data structure:
- Easy loading and saving to disk.
- Free memory on memory constraints.
- On-demand loading for optimal memory usage.
- Very fast insert and retrieve.
- Multi-thread-able and parallel-able usage.
- Pages should be linked together so you can do range queries
by going to the next page easily.
MGIndex takes the best features of a b+tree and improves upon
on them at the same time removing the impediments.
also extremely simple in design as the following diagram shows:
As you can see the page list is a sorted dictionary of first
keys from each page along with associated page number and page items
count. A page is a dictionary of key and record number pairs.
This format ensures a semi sorted key list, in that within a page the
data is not sorted but pages are in sort order relative to each other.
So a look-up for a key just compares the first keys in the page list to
find the page required and gets the key from the page's dictionary.
MGIndex is O(log M)+O(1), M being N / PageItemCount
PageItemCount = 10000 in the
Globals class]. This means that
you do a binary search in the page list in log M time and get the value
in O(1) time within a page.
RaptorDB starts off by loading the page list and it is good to go from there and pages are loaded on demand, based on usage.
In the event of page getting full and reaching the
MGIndex will sort the keys in the page's dictionary and split the data
in two pages ( similar to a b+tree split) and update the page list by
adding the new page and changing the first keys needed. This will ensure
the sorted page progression.
Interestingly the processor architecture plays an important role here
as you can see in the performance tests as it is directly related to
the sorting key time, the Core iX processors seem to be very good in
side effects of MGIndex
Here are some interesting side effects of
- Because the data pages are separate from the Page List
structure, implementing locking is easy and isolated within a
page and not the whole index, not so for normal trees.
- Splitting a page when full is simple and does not require a
tree traversal for node overflow checking as in a b+tree.
- Main page list updates are infrequent and hence the locking
of the main page list structure does not impact performance.
- The above make the
MGIndex a really good candidate for
The road not
taken / the road taken and doubled back!
Originally I used a
AATree found here (http://demakov.com/snippets/aatree.html) for the page
structures, for being extremely good and simple structure to
understand. After testing and comparing to the internal .net
(which is a Red-Black tree structure) it was
slower and so scrapped (see the performance comparisons).
I decided against using
SortedDictionary for the pages as it
was slower than a normal
Dictionary and for the purpose of a key
value store the sorted-ness was not need and could be handled in
other ways. You can switch to the
SortedDictionary in the code
at any time if you wish and it makes no difference to the
overall code other than you can remove the sorting in the page
I also tried an assorted number of sorting routines like double pivot
quick sort, timsort, insertion sort and found that they all were slower
than the internal .net quicksort routine in my tests.
In this version I have compiled a list of computers which I
have tested on and below is the results.
As you can see you get a very noticeable performance boost with the new Intel Core iX processors.
Comparing B+tree and MGIndex
For a measure of relative performance of a b+tree, Red/Black tree and
MGIndex I have compiled the following results.
Times are in seconds.
B+Tree : is the index code from
SortedDictionary : is the internal .net implementation which is said to be a Red/Black tree.
Really big data sets!
To really put the engine under pressure I did the following tests on huge data sets (times are in seconds, memory is in Gb) :
These tests were done on a HP ML120G6 system with 12Gb Ram, 10k raid
disk drives running Windows 2008 Server R2 64 bit. For a measure of
relative performance to
RaptorDb v1 I have included a 20 million test
with that engine also.
I deferred from testing the get test over 100 million record as it
would require a huge array in memory to store the
Guid keys for finding
later, that is why there is a NT (not tested) in the table.
Interestingly the read performance is relatively linear.
Index parameter tuning
To get the most out of
RaptorDB you can tune some parameters specific to your hardware.
PageItemCount : controls the size of each page.
Here are some of my results:
I have chosen the 10000 number as a good case in both read and writes,
you are welcome to tinker with this on your own systems and see what
works better for you.
Performance Tests v2.3
In v2.3 a single simple change of converting internal classes to structs rendered huge performance improvements of 2x+ and at least 30% lower memory usage. You are pretty much guaranteed to get 100k+ insert performance on any system.
Some of the test above were run 3 times because the computers were being used at the time (not cold booted for the tests) so the initial results were off. The HP G4 laptop is just astonishing.
I also re-ran the 100 million test on the last server in the above list and here is the results:
As you can see in the above test, the insert time is 4x faster (although the computer specs to not match the HP system tested earlier) and incredibly the memory usage is half than the previous test.
Using the Code
To create or open a database you use the following code :
var guiddb = RaptorDB.RaptorDB<Guid>.Open("c:\\RaptorDbTest\\multithread", false);
var strdb = RaptorDB.RaptorDB<string>.Open("c:\\intdb", 100, true);
To insert and retrieve data you use the following code :
Guid g = Guid.NewGuid();
if(guiddb.Get(g, out outstr))
The UnitTests project contains working example codes for different use cases so you can refer to it for more samples.
The following are a list of differences in v2 opposed to v1 of
- Log Files have been removed and are not needed anymore as
MGIndex is fast enough for in-process indexing.
- Threads have been replaced by timers.
- The index will be saved to disk in the background without
blocking the engine process.
- Messy generic code has been simplified and the need for a
has been removed, you can use normal int, long,
string and Guid data types.
RemoveKey has been added.
Other than that existing code should compile as is with the new
RaptorDBString and RaptorDBGuid
RaptorDBString is for long string keys (larger than 255
characters) and it is really useful for file paths etc. You can
use it in the following way :
var rap = new RaptorDBString(@"c:\raptordbtest\longstringkey", false);
var db = new RaptorDBGuid("c:\\RaptorDbTest\\hashedguid");
RaptorDBGuid is a special engine which will
MurMur2 hash the
Guid for lower memory usage (4 bytes opposed to 16 bytes),
this is useful if you have a huge number of items which you need
to store. You can use it in the following way :
var db = new RaptorDBGuid("c:\\RaptorDbTest\\hashedguid");
The following parameters are in the Global.cs file which you
can change which control the inner workings of the engine.
|Switch over point where
duplicates are stored as a WAH bitmap opposed to a list of
record numbers |
|The number of items within
a page |
|Background save index timer
seconds ( e.g. save the index to disk every 60 seconds) |
|Default string key size in
bytes (stored as UTF8) |
|Flush to storage file
|Compress and free bitmap
index memory on saves |
|Set Key and byte array
Value, returns void |
|Set Key and string Value,
returns void |
Get(T, out string)
|Get the Key and put it in
the string output parameter, returns true if key was
Get(T, out byte)
|Get the Key and put it in
the byte array output parameter, returns true if key was
|This will remove the key from the index|
|returns all the contents of
the main storage file as an |
KeyValuePair<T, byte> >
|Enumerate the Index from
the key given. |
|returns a list of main
storage file record numbers as an |
of the duplicate key specified
|returns the Value from the
main storage file as |
byte, used with
|returns the number of items
in the database index , counting the duplicates also if
|Allows the immediate save to
disk of the index (the engine will automatically save in
the background on a timer) |
|This will close all files
and stop the engine. |
In the event of a non clean shutdown
automatically rebuild the index from the last indexed item to
the last inserted item in the storage file. This feature also
enables you to delete the mgidx file and have
the index from scratch.
In v2 of
RaptorDB removing keys has been added with the following caveats :
- Data is not deleted from the storage file.
- A special delete record is added to the storage file for tracking deletes and which also help with index rebuilding when needed.
- Data is removed from the index.
The following unit tests are included in the source code (the output folder for all the tests is
- Duplicates_Set_and_Get : This test will generate 100 duplicates of 1000
Guids and fetch each one (This tests the WAH bitmap subsystem).
- Enumerate : This test will generate 100,001
Guids and enumerate the index from a predetermined
Guid and show the result count (the count will differ between runs).
- Multithread_test : This test will create 2 threads inserting 1,000,000 items and a third thread reading 2,000,000 items with a delay of 5 seconds from the start of insert.
- One_Million_Set_Get : This test will insert 1,000,000 items and read 1,000,000 items.
- One_Million_Set_Shutdown_Get : This test will do the above but shutdown and restart before reading.
- RaptorDBString_test : This test will create 100,000 1kb string keys and read them from the index.
- Ten_Million_Optimized_GUID : This test will use the
RaptorDBGuid class which will MurMur hash 10,000,000
Guids writting and reading them.
- Ten_Million_Set_Get : The same as 1 million test but with 10 million items.
- Twenty_Million_Optimized_GUID : The same as 10 million test but with 20 million items.
- Twenty_Million_Set_Get : The same as 1 million test but with 20 million items.
- StringKeyTest : A test for normal string keys of max 255 length.
- RemoveKeyTest : A test for removing keys works properly between shutdowns.
Format : *.mgdat
Values are stored in the following
structure on disk:
Format : *.mgbmp
Bitmap indexes are stored in the following format on disk :
The bitmap row is variable in length and will be reused if
the new data fits in the record size on disk, if not another
record will be created. For this reason a periodic index
compaction might be needed to remove unused records left
from previous updates.
Format : *.mgidx
The MGIndex index is saved in the following format as shown below:
Format : *.mgbmr , *.mgrec
Rec file is a series of
to disk with no special formatting.
These values map the record number to an offset in the
BITMAP index file and DOCS
- Initial Release v2.0 : 19th January 2012
- Update v2.1 : 26th January 2012
- lock on safedictionary iterator set, Thanks to igalk474
- string default(T) -> "" instead of null, Thanks to Ole Thrane for finding it
- mgindex string firstkey null fix
- added test for normal string keys
- fixed the link to the v1 article
- Update v2.2 : 8th February 2012
- bug fix removekey, Thanks to syro_pro
- removed un-needed initialization in safedictionary, Thanks to Paulo Zemek
- Update v2.3 : 1st March 2012
- changed internal classes to structs (2x+ speed, 30% less memory)
- added keystore class and code refactoring
- added a v2.3 performance section to the article
- Update v2.4 : 7th March 2012
- bug fix remove key set page isDirty -> Thanks to Martin van der Geer
- Page<T> is a class again to fix keeping it's state
- added RemoveKeyTest unit test
- removed MemoryStream from StorageFile.CreateRowHeader for speed
- current record number is also set in the bitmap index for duplicates
- Update v2.5 : 28th May 2012
- added SafeSortedList for access concurrency of the page list
- insert performance back to v2.3 speed (removed extra writing to duplicates)
- Update v2.6 : 20th December 2012
- post back code from RaptorDB the doc store
- added more data types (uint,short,double,float,datetime...)
- added locks to the indexfile
- updated logger
- updated safe dictionary with locks
- changed to Path.DirectorySeparatorChar for Mono/MonoDroid support
- bug fix edge case in WAHbitarray
- updated storage file
- Update v2.7.0 : 6th October 2013
- bug fix WAHBitArray
- index files are opened in shared mode for the ability of online copy backup
- dirty pages are sorted on save for read performance
- Update v2.7.5 : 11th October 2013
- bug fix saving page list to disk for counts > 50 million items