Change sync::rwlock to use atomics for read_count/read_mode instead of locking #7066
Labels
A-concurrency
Area: Concurrency
I-slow
Issue: Problems and improvements with respect to performance of generated code.
Comments
ghost
assigned
This was referenced Jun 13, 2013
bors
added a commit
that referenced
this issue
Jun 15, 2013
r? @brson links to issues: #7065 the race that's fixed; #7066 the perf improvement I added. There are also some minor cleanup commits here. To measure the performance improvement from replacing the exclusive with an atomic uint, I edited the ```msgsend-ring-rw-arcs``` bench test to do a ```write_downgrade``` instead of just a ```write```, so that it stressed the code paths that accessed ```read_count```. (At first I was still using ```write``` and saw no performance difference whatsoever, whoooops.) The bench test measures how long it takes to send 1,000,000 messages by using rwarcs to emulate pipes. I also measured the performance difference imposed by the fix to the ```access_lock``` race (which involves taking an extra semaphore in the ```cond.wait()``` path). The net result is that fixing the race imposes a 4% to 5% slowdown, but doing the atomic uint optimization gives a 6% to 8% speedup. Note that this speedup will be most visible in read- or downgrade-heavy workloads. If an RWARC's only users are writers, the optimization doesn't matter. All the same, I think this more than justifies the extra complexity I mentioned in #7066. The raw numbers are: ``` with xadd read count before write_cond fix 4.18 to 4.26 us/message with write_cond fix 4.35 to 4.39 us/message with exclusive read count before write_cond fix 4.41 to 4.47 us/message with write_cond fix 4.65 to 4.76 us/message ```
flip1995
pushed a commit
to flip1995/rust
that referenced
this issue
Apr 22, 2021
useless use of format! should return function directly fixes rust-lang#7066 changelog: [`useless_format`] wraps the content in the braces when it's needed. r? `@giraffate`
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The rwlocks are implemented following the algorithm described here: http://en.wikipedia.org/wiki/Readers-writers_problem#The_third_readers-writers_problem , which has a note:
Note that sections protected by counter_mutex could be replaced by a suitable fetch-and-add atomic instruction, saving two potential context switches in reader's code.
I should implement this, convince myself that it's right, and profile its performance against the old version.
You might argue that doing this the more clever way makes the code less maintainable, that it might accidentally have a bug that would be a lot harder to find. This is possible, although I might argue in response that the logic should never need to be changed, so if I get it right, it will be right forever. We'll have to see how much of a performance improvement it is.
The text was updated successfully, but these errors were encountered: