Synopse

Writing working multi-threaded code is not easy - it's even hard, as as a Delphi expert just wrote in his blog.

In fact, the first step into multi-thread application development could be:

"protect your shared variables with locks (aka critical sections), because you are not sure that the data you read/write is the same for all threads".

The CPU per-core cache is just one of the possible issues, which will lead into reading wrong values. Another issue which may lead into race condition is two threads writing to a resource at the same time: it's impossible to know which value will be stored afterward.

After a question on StackOverflow, I wanted to comment about the speed of generated code by diverse Delphi compiler versions.

Since performance matters when we write general purpose libraries like ours, we have some feedback to propose:

• The 4,500,000 unitary tests of our SQlite3 ORM Framework;
• The SynLZ compression algorithm.

For our ORM, we needed a class variable to be available for each TSQLRecord class type.

This variable is used to store the properties of this class type, i.e. the database Table properties (e.g. table and column names and types) associated with a particular TSQLRecord class, from which all our ORM objects inherit.

The class var statement was not enough for us:
- It's not available on earlier Delphi versions, and we try to have our framework work with Delphi 6-7 up to XE;
- This class var instance will be shared by all classes inheriting from the class where it is defined - and we need ONE instance PER class type, not ONE instance for ALL

We needed to find another way to implement this class variable. 

An unused VMT slot in the class type description was identified, then each class definition was patched in the process memory to contain our class variable.

How to make your software run fast, especially in a multi-threaded architecture?

We tried to remove the Memory Manager scaling problems in our SynScaleMM. It worked as expected in a multi-threaded server environment. Scaling is much better than FastMM4, for some critical tests. But it's not ready for production yet...

To be honest, the Memory Manager is perhaps not the bigger bottleneck in Multi-Threaded applications.

Here are some (not dogmatic, just from experiment and knowledge of low-level Delphi RTL) advice if you want to write FAST multi-threaded application in Delphi.

It was told on this Blog that the Delphi memory manager (FastMM4 since Borland 2006 - but it was even worse with the previous "Borland's" MM), doesn't scale well on multi-code CPU. That is, if you have a multi-threaded application with a lot of memory handling (e.g. aString := aString+someString), the Delphi MM won't scale with multi cores. When I mean "don't scale", I mean that the optimistic though of "my CPU has 4 cores, therefore the same work run in 4 threads will be 4 times faster than with 1 thread" is false. It performs even worse with multiple threads than with 1 thread...

So we went into forking a nice project, named ScaleMM, and created our scalable optimized MM, named SynScaleMM. Our forked modifications were even included in the main ScaleMM branch.

During my profiling of our SynScaleMM, I discovered some very nice results with Delphi compiler inlining features.

As I wrote in a previous post, Delphi string, dynamic array and memory manager don't like multi-core CPU.

My proposal is to add a threadlocalvar keyword, to be used instead of var in your code, to mark some variables to be used in only the current thread. Then the compiler and RTL won't have to use the LOCK instruction, and the application will be MUCH faster in multi-thread environment.

In a very interesting comment, PhiS spoke about the Mac OS X Stack Alignment problem, and the way asm code should be written for the future Cross Platform Delphi compiler. Here are some (hope without any Troll hidden) reflections I went through.

After some speed debates occurred in the Delphi community, I've rewritten the _CopyRecord function of the system.pas unit, with speed in mind.

This LGPL-distributed unit allows very fast JPEG image decoding, using SSE/SSE2 pure assembly code. It is much faster than the standard jpeg.pas unit.

It's based on the great Dr. Manhattan's source code, from http://sourceforge.net/projects/jpegdec