Performance and Effectiveness tests for v0.2.0

[GammaPi]

Benchmark

Parsec

Real Applications

Postgresql Mysql (Large DB) sqlite redis memcached mongodb apache nginx

[GammaPi]

The following problems in Coz may be discovered by Scaler:

• SQLite: Performance problem caused by indirect function call. The speedup is around 25.60%.
• fluidanimate: Both lines Coz identifies are in a custom barrier implementation, immediately before entering a loop that repeatedly calls pthread mutex trylock. 37.5% speedup.
• streamcluster: Call to a custom barrier slow. Replace pthread_barrier led to 68.4% speedup
• memcached: Remove a lock improves 9.39% speedup.

[GammaPi]

The following problems in Syncperf may be discovered by Scaler:

• Dedup: Asymmetric content problem, performance speedup is 12.1%
• Memcached-1.4.4: Single cache_lock to protect entire hash table.
• Mysql: get internal charset() 38.5%
• Facesim: One type of lock has 15288 acquisitions per second, but contention rate is very low. Solved by replace mytexlocks and conditional variables with atomic operations

[GammaPi]

SQLite: Performance problem caused by indirect function call. The speedup is around 25.60%. ❌

Coz discovered. Perf only report these account for 0.15% of total runtime. Confirmed through code analysis that Line 19616, Line 18974, Line 40345 are not calling other APIs.

[GammaPi]

fluidanimate, streamcluster: Both lines Coz identifies are in a custom barrier implementation, immediately before entering a loop that repeatedly calls pthread mutex trylock. 37.5% speedup.

Bug code found. This bug is possible to be found by Scaler.

Cannot reproduce the speedup reported by Coz ❌

[GammaPi]

Already finished previously.