June 15, 2006
Lock-free Interprocess CommunicationPassing Zero Words
Let us get back to the above question about the possibility of passing non-zero words. To solve this problem it is possible to use coding. For example, M arbitrary words are coded with M+1 non-zero words (M < 2^length(word)-1).
To perform coding of M arbitrary words into M+1 non-zero words we need to add one more word in front of these M words which will contain the number (from 1 to M) of the first zero word or a value bigger than M if there is no such zero-valued word inside those M words. Then the first zero-valued word is replaced with number of the next zero-valued word or with a number bigger than M if the next zero-valued word does not exist. The process repeats until all zero-valued words are replaced.
If we need to pass a chunk of M words to another process, it can be only one-write overhead in terms of the number of words written into shared memory. Some processor overhead is required for coding and decoding, but it should be relatively small because it is performed inside the same processor cache (no interprocess communication is involved). See Figure 4.
Figure 4: Passing zero words. White is for zero-valued words, black is for non-zero valued words, red for zero-valued words which were replaced with non-zero valued words and green is for the first additional word which contains the number of the first non-zero valued word.
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
Next Page
|
 |
|
|||||||||||||||||||||||||||||
|
|
|
