![Primitive Arrays
// Sequential Access = Predictable
for (int i=0; i<someArray.length; i++)
someArray[i]++;](https://image.slidesharecdn.com/performanceandpredictabilityrichardwarburton-141031061047-conversion-gate01/85/Performance-and-Predictability-Richard-Warburton-40-320.jpg)
![Primitive Arrays - Skipping Elements
// Holes Hurt
for (int i=0; i<someArray.length; i += SKIP)
someArray[i]++;](https://image.slidesharecdn.com/performanceandpredictabilityrichardwarburton-141031061047-conversion-gate01/85/Performance-and-Predictability-Richard-Warburton-41-320.jpg)
![Multidimensional Arrays
Multidimensional Arrays are really Arrays of
Arrays in Java. (Unlike C)
Some people realign their accesses:
for (int col=0; col<COLS; col++) {
for (int row=0; row<ROWS; row++) {
array[ROWS * col + row]++;
}
}](https://image.slidesharecdn.com/performanceandpredictabilityrichardwarburton-141031061047-conversion-gate01/85/Performance-and-Predictability-Richard-Warburton-43-320.jpg)
![Bad Access Alignment
Strides the wrong way, bad
locality.
array[COLS * row + col]++;
Strides the right way, good
locality.
array[ROWS * col + row]++;](https://image.slidesharecdn.com/performanceandpredictabilityrichardwarburton-141031061047-conversion-gate01/85/Performance-and-Predictability-Richard-Warburton-44-320.jpg)
Performance and Predictability - Richard Warburton
- 1. PERFORMANCE AND PREDICTABILITY Richard Warburton @richardwarburto insightfullogic.com
- 3. Why care about low level rubbish? Branch Prediction Memory Access Storage Conclusions
- 4. Technology or Principles “60 messages per second.” “8 ElasticSearch servers on AWS, 26 front end proxy serves. Double that in backend app servers.” 60 / (8 + 26 + 2 * 26) = 0.7 messages / server / second. http://highscalability.com/blog/2014/1/6/how-hipchat-stores-and-indexes-billions-of-messages-using- el.html
- 5. “What we need is less hipsters and more geeks” - Martin Thompson
- 6. Performance Discussion Product Solutions “Just use our library/tool/framework, and everything is web-scale!” Architecture Advocacy “Always design your software like this.” Methodology & Fundamentals “Here are some principles and knowledge, use your brain“
- 8. Case Study: Messaging ● 1 Thread reading network data ● 1 Thread writing network data ● 1 Thread conducting admin tasks
- 9. Unifying Theme: Be Predictable An opportunity for an underlying system: ○ Branch Prediction ○ Memory Access ○ Hard Disks
- 10. Do you care? Many problems not Predictability Related Networking Database or External Service Minimising I/O Garbage Collection Insufficient Parallelism Use an Optimisation Omen
- 11. Why care about low level rubbish? Branch Prediction Memory Access Storage Conclusions
- 12. What 4 things do CPUs actually do?
- 13. Fetch, Decode, Execute, Writeback
- 14. Pipelined
- 17. What about branches? public static int simple(int x, int y, int z) { int ret; if (x > 5) { ret = y + z; } else { ret = y; } return ret; }
- 18. Branches cause stalls, stalls kill performance
- 19. Can we eliminate branches?
- 20. Strategy: predict branches and speculatively execute
- 21. Static Prediction A forward branch defaults to not taken A backward branch defaults to taken
- 23. Conditional Branches if(x == 0) { x = 1; } x++; mov eax, $x cmp eax, 0 jne end mov eax, 1 end: inc eax mov $x, eax
- 24. Static Hints (Pentium 4 or later) __emit 0x3E defaults to taken __emit 0x2E defaults to not taken don’t use them, flip the branch
- 25. Dynamic prediction: record history and predict future
- 26. Branch Target Buffer (BTB) a log of the history of each branch also stores the program counter address its finite!
- 27. Local record per conditional branch histories Global record shared history of conditional jumps
- 28. Loop specialised predictor when there’s a loop (jumping in a cycle n times) Function specialised buffer for predicted nearby function returns N level Adaptive Predictor accounts for up patterns of up to N+1 if statements
- 29. Optimisation Omen Use Performance Event Counters (Model Specific Registers) Can be configured to store branch prediction information Profilers & Tooling: perf (linux), VTune, AMD Code Analyst, Visual Studio, Oracle Performance Studio
- 30. Demo perf
- 31. Summary CPUs are Super-pipelined and Superscalar Branches cause stalls Simplify your code! Especially branching logic and megamorphic callsites
- 32. Why care about low level rubbish? Branch Prediction Memory Access Storage Conclusions
- 33. The Problem Very Fast Relatively Slow
- 34. The Solution: CPU Cache Core Demands Data, looks at its cache If present (a "hit") then data returned to register If absent (a "miss") then data looked up from memory and stored in the cache Fast memory is expensive, a small amount is affordable
- 35. Multilevel Cache: Intel Sandybridge Physical Core 0 HT: 2 Logical Cores Level 1 Instruction Cache Shared Level 3 Cache Level 1 Data Cache Level 2 Cache .... Physical Core N HT: 2 Logical Cores Level 1 Data Cache Level 1 Instruction Cache Level 2 Cache
- 36. How bad is a miss? Location Latency in Clockcycles Register 0 L1 Cache 3 L2 Cache 9 L3 Cache 21 Main Memory 150-400
- 37. Prefetching Eagerly load data Adjacent & Streaming Prefetches Arrange Data so accesses are predictable
- 38. Temporal Locality Repeatedly referring to same data in a short time span Spatial Locality Referring to data that is close together in memory Sequential Locality Referring to data that is arranged linearly in memory
- 39. General Principles Use smaller data types (-XX:+UseCompressedOops) Avoid 'big holes' in your data Make accesses as linear as possible
- 40. Primitive Arrays // Sequential Access = Predictable for (int i=0; i<someArray.length; i++) someArray[i]++;
- 41. Primitive Arrays - Skipping Elements // Holes Hurt for (int i=0; i<someArray.length; i += SKIP) someArray[i]++;
- 42. Primitive Arrays - Skipping Elements
- 43. Multidimensional Arrays Multidimensional Arrays are really Arrays of Arrays in Java. (Unlike C) Some people realign their accesses: for (int col=0; col<COLS; col++) { for (int row=0; row<ROWS; row++) { array[ROWS * col + row]++; } }
- 44. Bad Access Alignment Strides the wrong way, bad locality. array[COLS * row + col]++; Strides the right way, good locality. array[ROWS * col + row]++;
- 45. Full Random Access L1D - 5 clocks L2 - 37 clocks Memory - 280 clocks Sequential Access L1D - 5 clocks L2 - 14 clocks Memory - 28 clocks
- 46. Data Layout Principles Primitive Collections (GNU Trove, GS-Coll, FastUtil, HPPC) Arrays > Linked Lists Hashtable > Search Tree Avoid Code bloating (Loop Unrolling)
- 47. Custom Data Structures Judy Arrays an associative array/map kD-Trees generalised Binary Space Partitioning Z-Order Curve multidimensional data in one dimension
- 48. Data Locality vs Java Heap Layout 0 1 2 class Foo { Integer count; Bar bar; Baz baz; } // No alignment guarantees for (Foo foo : foos) { foo.count = 5; foo.bar.visit(); } 3 ... Foo count bar baz
- 49. Data Locality vs Java Heap Layout Serious Java Weakness Location of objects in memory hard to guarantee. GC also interferes Copying Compaction
- 50. Optimisation Omen Again Use Performance Event Counters Measure for cache hit/miss rates Correlate with Pipeline Stalls to identify where this is relevant
- 51. Object Layout Control On Heap http://objectlayout.github.io/ObjectLayout Off Heap - Data Structures: Chronicle or JCTools Experimental - Serialisation: SBE, Cap’n’p, Flatbuffers
- 52. Summary Cache misses cause stalls, which kill performance Measurable via Performance Event Counters Common Techniques for optimizing code
- 53. Why care about low level rubbish? Branch Prediction Memory Access Storage Conclusions
- 54. Hard Disks Commonly used persistent storage Spinning Rust, with a head to read/write Constant Angular Velocity - rotations per minute stays constant Sectors size differs between device
- 55. A simple model Zone Constant Angular Velocity (ZCAV) / Zoned Bit Recording (ZBR) Operation Time = Time to process the command Time to seek Rotational speed latency Sequential Transfer TIme
- 56. ZBR implies faster transfer at limits than centre (~25%)
- 57. Seeking vs Sequential reads Seek and Rotation times dominate on small values of data Random writes of 4kb can be 300 times slower than theoretical max data transfer Consider the impact of context switching between applications or threads
- 58. Fragmentation causes unnecessary seeks
- 62. Optimisation Omen 1. Application Spending time waiting on I/O 2. I/O Subsystem not transferring much data
- 63. Summary Simple, sequential access patterns win Fragmentation is your enemy Alignment can be important
- 64. Why care about low level rubbish? Branch Prediction Memory Access Storage Conclusions
- 65. Speedups ● Possible 20 cycle stall for a mispredict (example 5x slowdown) ● 200x for L1 cache hit vs Main Memory ● 300x for sequential vs random on disk ● Theoretical Max
- 66. Latency Numbers L1 cache reference 0.5 ns Branch mispredict 5 ns L2 cache reference 7 ns 14x L1 cache Mutex lock/unlock 25 ns Main memory reference 100 ns 20x L2 cache, 200x L1 cache Compress 1K bytes with Zippy 3,000 ns Send 1K bytes over 1 Gbps network 10,000 ns 0.01 ms Read 4K randomly from SSD* 150,000 ns 0.15 ms Read 1 MB sequentially from memory 250,000 ns 0.25 ms Round trip within same datacenter 500,000 ns 0.5 ms Read 1 MB sequentially from SSD* 1,000,000 ns 1 ms Disk seek 10,000,000 ns 10 ms Read 1 MB sequentially from disk 20,000,000 ns 20 ms Send packet CA->Netherlands->CA 150,000,000 ns 150 ms Stolen (cited) from https://gist.github.com/jboner/2841832
- 67. Common Themes ● Principles over Tools ● Data over Unsubstantiated Claims ● Simple over Complex ● Predictable Access over Random Access
- 68. More information Articles http://www.akkadia.org/drepper/cpumemory.pdf https://gmplib.org/~tege/x86-timing.pdf http://psy-lob-saw.blogspot.co.uk/ http://www.intel.com/content/www/us/en/architecture-and-technology/64- ia-32-architectures-optimization-manual.html http://mechanical-sympathy.blogspot.co.uk http://www.agner.org/optimize/microarchitecture.pdf Mailing Lists: https://groups.google.com/forum/#!forum/mechanical-sympathy https://groups.google.com/a/jclarity.com/forum/#!forum/friends http://gee.cs.oswego.edu/dl/concurrency-interest/
- 70. Q & A @richardwarburto insightfullogic.com tinyurl.com/java8lambdas