June 15, 2006
Lock-free Interprocess CommunicationPerformance Testing
Algorithm 2 was compared to Algorithm 3 and a standard version of this algorithm. This test involves passing 10,000,000 integer values from one thread to another. The standard algorithm performs all required locking when it accesses shared resources. Locking is performed for every passing integer.
Table 1: Performance comparison of standard algorithm, Algorithm 2 and Algorithm 3.
I would not consider this comparison very fair for standard applications because we usually do not send information in small chunks from one process to another, but for some applications (like routers or switches) this technique may bring some benefits. For example, consider a few specialized processors which perform pipelined data processing (Figure 6).
Figure 6: Pipelined data processing.
Then data can be passed between these processors using the proposed mechanism. It should improve overall performance in case of high load and probably decrease latency time in case of low load.
Table 2 shows Algorithm 4 compared with standard version of this algorithm. The testing procedure in this case involves running 8 threads. Each thread reads from the registry 1,024,000 times and updates the registry 1,000 times.
Table 2: Performance comparison of standard and optimized versions of Algorithm 4.
The lock-free algorithm does not provide any benefits on single-processor systems, but on a multi-processor system it outperformed the standard algorithm by about a factor of 30. Since multiprocessor systems are becoming mainstream the described technique or its variations may bring considerable benefits for applications which use it.
References:
Previous Page |
1 Lock-free IPC
|
2 Algorithm One
|
3 Algorithm Two (Light Pipe)
|
4 Passing Zero Words
|
5 Algorithm Three (Cache Line Optimized Light Pipe)
|
6 Algorithm Four (Read-optimized Registry)
|
7 Performance Testing
|
 |
|
||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|
