This presentation uses the following non-standard fonts: Lato - PowerPoint PPT Presentation

Progress Points Little’s Law: W= L / λ latency = transactions / throughput � � � � � � � � � transactions++ transactions-- Bob wants to minimize response Bme. He adds latency progress points.

Coz: a Causal Profiler for Linux > coz run --- ./ogle_server deploy

Coz Produces Causal Profiles Let’s use it to improve Ogle � Program Speedup � � � � Speedup ?

Using Causal Profiling on Ogle gle gle found 8,000,000 
 similar images

Using Causal Profiling on Ogle dedup compression ferret image comparison

Ferret image comparison input output feature segmentaOon indexing ranking extracOon

Ferret 100% ● ● 75% ● ● ● line 320 ● ● ● ranking 50% ● ● ● ● ● ● 25% ● ● ● ● ● 0% ● 100% Program Speedup 75% line 358 indexing 50% 25% 0% 100% 75% line 255 segmentaOon 50% 25% 0% 0% 25% 50% 75% 100% Line Speedup ● line 320 Line line 358 line 255

Ferret input output feature segmentaOon indexing ranking extracOon

Ferret Probably doesn’t need as many threads input output feature segmentaOon indexing ranking extracOon

Ferret 21% Speedup input output feature segmentaOon indexing ranking extracOon

What did Causal Profiling predict? ranking Increased from 16 to 22 threads 27% increase in ranking throughput

What did Causal Profiling predict? 100% ● ● 75% ● ● ● line 320 ● ● ● ranking 50% ● ● ● ● ● ● 25% ● ● ● ● ● 0% 0% ● 100% 0% 25% 50% 75% 100% Program Speedup Line Speedup 27% increase in ranking throughput Causal Profiling predicted a 21% improvement Exactly what we observed

Using Causal Profiling on Ogle dedup compression ferret image comparison

Dedup Compression via deduplicaBon

Dedup Compression via deduplicaBon

Dedup Compression via deduplicaBon grumpycat1.jpg funisawful.jpg

Dedup Compression via deduplicaBon i = hash_function( )

Dedup Compression via deduplicaBon i = hash_function( )

Dedup Compression via deduplicaBon i = hash_function( )

Dedup Compression via deduplicaBon i = hash_function( )

Dedup Compression via deduplicaBon i = hash_function( )

Dedup Compression via deduplicaBon i = hash_function( )

Dedup Compression via deduplicaBon Hash table is accessed concurrently by many threads � � � � � � Causal Profiler says the loop that accesses this list is important

Dedup Compression via deduplicaBon More hash buckets should lead to fewer collisions � � � � � � � � � � � �

Dedup Compression via deduplicaBon More hash buckets should lead to fewer collisions � � � � � � � � � � � �

Dedup Compression via deduplicaBon More hash buckets should lead to fewer collisions No performance improvement � � � � � � � � � � � �

Dedup Compression via deduplicaBon What else could be causing collisions? i = hash_function( ) � � � � � � � � � � � �

Dedup Compression via deduplicaBon Horrible hash func8on! Items per-bucket Bin UBlizaBon 150 Original Version 2.3% 100 50 Collisions 0 0 250 500 750 1000 Bucket Index

Dedup Compression via deduplicaBon Bin UBlizaBon 150 Original Version 2.3% 100 Items per-bucket 50 Collisions Collisions 0 150 After Optimization 9% Speedup 100 82% 50 0 0 250 500 750 1000 Bucket Index

Dedup Compression via deduplicaBon What did Causal Profiling predict? Blocks per-bucket Before: 76.7 ACer: 2.09 96% traversal speedup 9% predicted speedup, 
 exactly what we observed

Using Causal Profiling on Ogle dedup compression ferret image comparison

Simple SQL Database #if THREAD_SAFE config_t global_config = { … .unlock = pthread_mutex_unlock, .getsize = sqlite_usable_size, .nextitem = sqlite_pagecache_next, … }; #endif

Simple SQL Database void pthreadMutexLeave(lock* l) { global_config.unlock(l); } Indirect Call .nextitem = sqlite_pagecache_next, Cheap, but almost the same cost as pthread_mutex_unlock

Simple SQL Database Coz highlights these lines void pthreadMutexLeave(lock* l) { global_config.unlock(l); } void sqlite3MemSize(void* p) { global_config.getsize(p); } void pcache1Fetch(item* i) { global_config.nextitem(i); }

Simple SQL Database Line 16916 Line 18974 Line 40345 25% ● ● ● ● ● ● ● 0% Program Speedup ● ● ● ● ● ● ● − 25% ● − 50% 0% 50% 0% 50% 0% 50% Line Speedup