1. Building Multithreaded Solutions with OmniThreadLibrary
The only URL you have to write down:
Primož Gabrijelčič,

2. Ask how was at the previous presentation and maybe replay few slides from it (stuff about OTL)
OmniThreadLibrary

3. OmniThreadLibrary Multithreading library for Delphi 2007/2009⇨
Task oriented
Strong messaging support
Can be used with any threading infrastructure
High level parallel programming

4. Project Status Free “As in Air” -Theo de Raadt Available
OpenBSD license
Available
code.google.com/p/omnithreadlibrary/
Actively developed
830 commits
Used
downloads of the latest release
Almost no documentation
otl.17slon.com/tutorials.htm
830 commits in 2 years 3 months
1776 last time I looked; 1.05a was released in March
Plethora of examples
Almost no documentation 

5. Today’s Topics Communication vs. Shared data Tasks vs. Threads
Thread pools

6. I’ll show the communication-related code later, together with task-related code
Tasks

7. Task <> Thread Task is part of code that has to be executed
Thread is execution environment
You take care of the task, OTL takes care of the thread

8. Execution Models CreateTask(task_procedure) CreateTask(task_method )
CreateTask(TOmniWorker_object)
CreateTask(anonymous_procedure)

9. Just a short info on thread pools before I open the Delphi …

10. Thread pool Starting up a thread takes time
Thread pool keeps threads alive and waits for tasks
Automatic thread startup/shutdown
User code executed at thread creation
Connection pool
.Run ⇨ .Schedule
Automatic – according to parameters specified by the programmer (min/max threads, max idle time)

11. Communication

12. Shared Data Pros Cons Keep in mind Only one copy Fast if only reading
Locking
Bad scaling
Deadlocks, livelocks
Keep in mind
Fine-grained is better
Optimistic locking
Fast if not modified: copy on write

13. Messaging Pros Cons No shared data - no* locking Hard to understand
Increased memory usage
*For sufficiently flexible definition of “No”
Hard to understand = sometimes
Increased memory usage = sometimes

14. Common Sense Sometimes you need both Minimize shared data
Minimize interaction points
“If your solution depends on sharing data million times a second, you’re doomed.” -me
And it doesn’t matter if you’re locking million times a second or sending million messages a second.

15. Messaging Solutions Windows messages Pipes Mailslots Sockets (TCP/IP)
Shared memory + (micro)locking
Used by the OmniThreadLibrary communication primitives
“shared memory” in context of “shared data inside application” not “memory shared between processes”

16. Messaging in the OTL Bounded stack Bounded queue Dynamic queue
Dynamically allocated, O(1) enqueue and dequeue
Useful for single-threaded applications too
All are threadsafe, microlocking, and support multiple readers and writers
All are lock-free (or micro-locking), all are threadsafe, all support multiple readers and writers
Bounded = size specified at creation

17. TOmniValue Used in OTL messaging Record with operators
Can contain strings, interfaces, objects …
Faster than Variant
Much faster than TValue (but slightly less powerfull)

18. Show me the code!

19. Danger, Will Robinson!
“New programmers are drawn to multithreading like moths to flame, with similar results.”
-Danny Thorpe
Image © Hammacher Schlemmer

20. Be Afraid Be very afraid! Designing parallel solutions is hard
Writing multithreaded code is hard
Testing multicore apps is hard
Debugging multithreading code is pure insanity
Yes, next two slides are copied from the previous presentation. It’s just that I can’t repeat this often enough …

21. Keep in Mind Don’t parallelize everything
Don’t create thousands of threads
Rethink the algorithm
Prove the improvements
Test, test and test

22. Q & A