Synopse Open Source

Insertion speed

Here we insert 5,000 rows of data, with diverse scenarios:

• 'Direct' stands for a individual Client.Add() insertion;
• 'Batch' mode, has already been described in this blog;
• 'Trans' indicates that all insertion is nested within a transaction - which makes a great difference, e.g. with a SQlite3 database.

Benchmark was run on a Core i7 notebook, with standard SSD, including anti-virus and background applications, over a 100 Mb corporate network, linked to a shared Oracle 11g database. A local instance of MSSQLExpress 2008 R2 was running locally. So it was a development environment, very similar to low-cost production site, not dedicated to give best performance. During the process, CPU was noticeable used only for SQLite3 in-memory and TObjectList - most of the time, the bottleneck is not the CPU, but the storage or network. As a result, rates and timing may vary depending on network and server load, but you get results similar to what could be expected on customer side, with an average hardware configuration.

Due to its ACID implementation, SQLite3 process on file waits for the hard-disk to have finished flushing its data, therefore it is the reason why it is slower than other engines at individual row insertion (less than 10 objects per second with a mechanical hardrive instead of a SDD) outside the scope of a transaction.

So if you want to reach the best writing performance in your application with the default engine, you should better use transactions and regroup all writing into services or a BATCH process. Another possibility could be to execute DB.Synchronous := smOff and/or DB.LockingMode := lmExclusive at SQLite3 engine level before process: in case of power loss at wrong time it may corrupt the database file, but it will increase the rate by a factor of 50 (with hard drive), as stated by the "off" and "off exc" rows of the table. Note that by default, the FireDAC library set both options, so results above are to be compared with "SQLite3 off exc" rows.

For both our direct Oracle access SynDBOracle.pas library and FireDAC, Batch process benefit of the array binding feature a lot (known as Array DML in FireDAC/AnyDAC).

Reading speed

Now the same data is retrieved via the ORM layer:

• 'By one' states that one object is read per call (ORM generates a SELECT * FROM table WHERE ID=? for Client.Retrieve() method);
• 'All *' is when all 5000 objects are read in a single call (i.e. running SELECT * FROM table from a FillPrepare() method call), either forced to use the virtual table layer, or with direct static call.

Here are some reading speed values, in objects/second:

The SQLite3 layer gives amazing reading results, which makes it a perfect fit for most typical ORM use. When running with DB.LockingMode := lmExclusive defined (i.e. "off exc" rows), reading speed is very high, and benefits from exclusive access to the database file. External database access is only required when data is expected to be shared with other processes.

In the above table, it appears that all libraries based on DB.pas are slower than the others for reading speed. In fact, TDataSet sounds to be a real bottleneck, due to its internal data marshalling. Even FireDAC, which is known to be very optimized for speed, is limited by the TDataSet structure. Our direct classes, or even ZEOS/ZDBC performs better, since they are able to output JSON content with no additional marshalling.

For both writing and reading, TObjectList / TSQLRestServerStaticInMemory engine gives impressive results, but has the weakness of being in-memory, so it is not ACID by design, and the data has to fit in memory. Note that indexes are available for IDs and stored AS_UNIQUE properties.

As a consequence, search of non-unique values may be slow: the engine has to loop through all rows of data. But for unique values (defined as stored AS_UNIQUE), both insertion and search speed is awesome, due to its optimized O(1) hash algorithm - see the following benchmark, especially the "By name" row for "TObjectList" columns, which correspond to a search of an unique RawUTF8 property value via this hashing method.

Above table results were run on a Core 2 duo laptop, so numbers are lower than with the previous tables.

Analysis and use case proposal

When declared as virtual table (via a VirtualTableRegister call), you have the full power of SQL (including JOINs) at hand, with incredibly fast CRUD operations: 100,000 requests per second for objects read and write, including serialization and Client-Server communication!

Some providers are first-class citizens to mORMot, like SQLite3, Oracle, or MS SQL. You can connect to them without the bottleneck of the DB.pas unit, nor any restriction of your Delphi license (a Starter edition is enough). For instance, SQLite3 could be used as main database engine for a client-server application with heavy concurrent access - if you have doubts about its scaling abilities, see this blog article. Direct access to Oracle is also available, with impressive results in BATCH mode (aka array binding). MS SQL Server, directly accessed via OleDB (or ODBC) gives pretty good timing, and a MS SQL Server 2008 R2 Express instance is a convincing option, for a very offerdable price (i.e. for free) - the LocalDB (MSI installer) edition is enough to start with. Any other OleDB, ODBC or ZDBC providers may also be used, with direct access. For instance, Firebird embedded gives pretty consistent timing, when accessed via Zeos/ZDBC.

But mORMot is very open-minded: you can use any DB.pas provider, e.g. FireDAC, UniDAC, DBExpress, NexusDB or even the BDE, but with the additional layer introduced by using a TDataSet instance, at reading.

Note that all those tests were performed locally and in-process, via a TSQLRestClientDB instance. For both insertion and reading, a Client-Server architecture (e.g. using HTTP/1.1 for mORMot clients) will give even better results for BATCH and retrieve all modes.

During the tests, internal caching was disabled, so you may expect speed enhancements for real applications, when data is more read than written: for instance, when an object is retrieved from the cache, you achieve more than 700,000 read requests per second, whatever database is used.

Therefore, the typical use may be the following:

Whatever database back-end is used, don't forget that mORMot design will allow you to switch from one library to another, just by changing a TSQLDBConnectionProperties class type. And note that you can mix external engines, on purpose: you are not tied to one single engine, but the database access can be tuned for each ORM table, according to your project needs.