You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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
```
0 commit comments