To test the code, I wrote a monstrous “basher.” It ran lots of transactions, each of which contained nested transactions, recursively up to some maximum nesting depth. Each of the nested transactions would lock and read several elements of a shared array in ascending order and add something to each element, preserving the invariant that the sum of all the elements in the array was zero. Each subtransaction was either committed or aborted—90 percent commits, 10 percent aborts, or whatever. Multiple threads ran these transactions concurrently and beat on the array for a prolonged period. Since it was a shared-memory facility that I was testing, I ran multiple multithreaded bashers concurrently, each in its own process.
At reasonable concurrency levels, the basher passed with flying colors. But when I really cranked up the concurrency, I found that occasionally, just occasionally, the basher would fail its consistency check. I had no idea what was going on. Of course I assumed it was my fault because I had written all of this new code.
......
I spent a week or so writing painfully thorough unit tests of each component, and all the tests passed. Then I wrote detailed consistency checks for each internal data structure, so I could call the consistency checks after every mutation until a test failed. Finally I caught a low-level consistency check failing—not repeatably, but in a way that allowed me to analyze what was going on. And I came to the inescapable conclusion that my locks weren’t working. I had concurrent read-modify-write sequences taking place in which two transactions locked, read, and wrote the same value and the last write was clobbering the first.
I had written my own lock manager, so of course I suspected it. But the lock manager was passing its unit tests with flying colors. In the end, I determined that what was broken wasn’t the lock manager, but the underlying mutex implementation! This was before the days when operating systems supported threads, so we had to write our own threading package. It turned out that the engineer responsible for the mutex code had accidentally exchanged the labels on the lock and try-lock routines in the assembly code for our Solaris threading implementation. So every time you thought you were calling lock, you were actually calling try-lock, and vice versa. Which means that when there was actual contention—rare in those days—the second thread just sailed into the critical section as if the first thread didn’t have the lock. The funny thing was that that this meant the whole company had been running without mutexes for a couple weeks, and nobody noticed.
[...] learning parallel programming is no harder than learning a second programming language.Let me be clear: I am not violently disagreeing with Mr. Breshears, but let me put it this way: I would suggest that for a reasonably intelligent computer science student, learning a second programming language, on a scale of 1-10, is around a "3" (in terms of difficulty). Now, for this same student to learn parallel programming, I would estimate that, on a scale of 1-10, the amount of difficulty would be a "7".
