This paper describes the bakery algorithm for implementing mutual exclusion. I have invented many concurrent algorithms. I feel that I did not invent the bakery algorithm, I discovered it. Like all shared-memory synchronization algorithms, the bakery algorithm requires that one process be able to read a word of memory while another process is writing it. (Each memory location is written by only one process, so concurrent writing never occurs.) Unlike any previous algorithm, and almost all subsequent algorithms, the bakery algorithm works regardless of what value is obtained by a read that overlaps a write. If the write changes the value from 0 to 1, a concurrent read could obtain the value 7456 (assuming that 7456 is a value that could be in the memory location). The algorithm still works. I didn’t try to devise an algorithm with this property. I discovered that the bakery algorithm had this property after writing a proof of its correctness and noticing that the proof did not depend on what value is returned by a read that overlaps a write.<\/p>\n
I don’t know how many people realize how remarkable this algorithm is. Perhaps the person who realized it better than anyone is Anatol Holt, a former colleague at Massachusetts Computer Associates. When I showed him the algorithm and its proof and pointed out its amazing property, he was shocked. He refused to believe it could be true. He could find nothing wrong with my proof, but he was certain there must be a flaw. He left that night determined to find it. I don’t know when he finally reconciled himself to the algorithm’s correctness.<\/p>\n
Several books have included emasculated versions of the algorithm in which reading and writing are atomic operations, and called those versions “the bakery algorithm”. I find that deplorable. There’s nothing wrong with publishing a simplified version, as long as it’s called a simplified version.<\/p>\n
What is significant about the bakery algorithm is that it implements mutual exclusion without relying on any lower-level mutual exclusion. Assuming that reads and writes of a memory location are atomic actions, as previous mutual exclusion algorithms had done, is tantamount to assuming mutually exclusive access to the location. So a mutual exclusion algorithm that assumes atomic reads and writes is assuming lower-level mutual exclusion. Such an algorithm cannot really be said to solve the mutual exclusion problem. Before the bakery algorithm, people believed that the mutual exclusion problem was unsolvable–that you could implement mutual exclusion only by using lower-level mutual exclusion. Brinch Hansen said exactly this in a 1972 paper. Many people apparently still believe it. (See [91]<\/a>.)<\/p>\n The paper itself does not state that it is a “true” mutual exclusion algorithm. This suggests that I didn’t realize the full significance of the algorithm until later, but I don’t remember.<\/p>\n For a couple of years after my discovery of the bakery algorithm, everything I learned about concurrency came from studying it. Papers like [25]<\/a>, [33]<\/a>, and [70]<\/a> were direct results of that study. The bakery algorithm was also where I introduced the idea of variables belonging to a process–that is, variables that could be read by multiple processes, but written by only a single process. I was aware from the beginning that such algorithms had simple distributed implementations, where the variable resides at the owning process, and other processes read it by sending messages to the owner. Thus, the bakery algorithm marked the beginning of my study of distributed algorithms.<\/p>\n