Products – NOBLE – Performance
 
NOBLE Professional Edition is designed using the very latest non-blocking techniques for achieving best possible performance and scalability. The actual performance in real applications naturally depends on the specific run-time environment and application design. Therefore, NOBLE Professional Edition contains several different versions of each data structure so that the developer can choose the best version for the specific conditions.

Simple non-blocking techniques have been used in a large scope of commercial applications since several years. These applications span from business and trade applications, music and synthesizer applications to even operating system kernels. However, more advanced data structures using non-blocking technology haven't been available for developers until now. With help from NOBLE Professional Edition can developers now use non-blocking versions of a large and rich set of data structures based on state-of-the-art algorithmic solutions.

Benchmarks

Here follows some examples of stress-tests on actual platforms, i.e. using the Core i7 920 processor from Intel containing four processor cores that turn up as eight logical processors, thus supporting real parallelism. As can be seen, the various implementations of a certain component have quite distinct performance characteristics depending on internal as well as environmental properties, thus illustrating the usefulness of easily supporting run-time selection of actual implementation as offered by NOBLE Professional Edition. 

The above graph shows an experiment with the different versions of the Queue abstract data type currently available in NOBLE Professional Edition. The threads are accessing the Queue with a medium load; between every operation on the Queue, 500 floating-point arithmetic operations are performed.

The above graph shows an experiment where the threads are accessing the Queue with a maximum load; i.e. the threads are only performing operations on the Queue.

The above graph shows an experiment with several different versions of the Dictionary (also known as Map) abstract data type, running with an average size of 1000 associated items. The threads are accessing the Dictionary with a medium load; between every operation on the Dictionary, 500 floating-point arithmetic operations are performed.

The above graph shows an experiment where the threads are accessing the Dictionary with a maximum load; i.e. the threads are only performing operations on the Dictionary.

For more relevant benchmarks showing the advantages in performance for non-blocking techniques compared to traditional techniques (e.g. lock-based using mutexes or semaphores), please see our white papers available in the technical documents section. Please note that besides increased parallelism, non-blocking technology also offers other very important features as increased stability and termination guarantees, e.g. threads can be killed even during an operation on a data structure without affecting the stability of the other threads.

Note that the differences in performance and scalability behavior between non-blocking synchronization and traditional techniques will be even more significant with increasing number of cores, processors and parallelism.