Davidlohr Bueso <dave@xxxxxxxxxxxx> writes:
I ran this on a 40-core (no ht) Westmere with two benchmarks. The first
is Manfred's sysvsem lockunlock[1] program which uses _processes_ to,
well, lock and unlock the semaphore. The options are a little
unconventional, to keep the "critical region small" and the lock+unlock
frequency high I added busy_in=busy_out=10. Similarly, to get the
worst case scenario and have everyone update the same semaphore, a single
one is used. Here are the results (pretty low stddev from run to run)
for doing 100,000 lock+unlock.
- 1 proc:
* vanilla
total execution time: 0.110638 seconds for 100000 loops
* dirty
total execution time: 0.120144 seconds for 100000 loops
- 2 proc:
* vanilla
total execution time: 0.379756 seconds for 100000 loops
* dirty
total execution time: 0.477778 seconds for 100000 loops
- 4 proc:
* vanilla
total execution time: 6.749710 seconds for 100000 loops
* dirty
total execution time: 4.651872 seconds for 100000 loops
- 8 proc:
* vanilla
total execution time: 5.558404 seconds for 100000 loops
* dirty
total execution time: 7.143329 seconds for 100000 loops
- 16 proc:
* vanilla
total execution time: 9.016398 seconds for 100000 loops
* dirty
total execution time: 9.412055 seconds for 100000 loops
- 32 proc:
* vanilla
total execution time: 9.694451 seconds for 100000 loops
* dirty
total execution time: 9.990451 seconds for 100000 loops
- 64 proc:
* vanilla
total execution time: 9.844984 seconds for 100032 loops
* dirty
total execution time: 10.016464 seconds for 100032 loops
Lower task counts show pretty massive performance hits of ~9%, ~25%
and ~30% for single, two and four/eight processes. As more are added
I guess the overhead tends to disappear as for one you have a lot
more locking contention going on.
Can you check your notes on the 4 process case? As I read the 4 process
case above it is ~30% improvement. Either that is a typo or there is the
potential for quite a bit of noise in the test case.