Java Performance Engineer's Survival Guide
14 likes3,700 views
The document focuses on performance engineering, emphasizing the importance of understanding quality of service (QoS) and setting measurable service level agreements (SLAs) to enhance user satisfaction. It details methods for monitoring, measuring, and optimizing throughput, response time, and capacity within the application performance landscape. The text also outlines top-down and bottom-up approaches for analyzing and improving system performance, highlighting the need for appropriate tools and techniques in both strategies.
Java Performance Engineer's Survival Guide
- 2. About Me
- 7. AskQuestions ©2016 CodeKaram6 • What makes the users happy? • What makes them unhappy? • Any current issues that need to be tackled?
- 8. Define and Understand ©2016 CodeKaram7 • Understand QoS (Quality of Service) for end users • Define QoS success metrics in measurableterms • Those are the service level agreements (SLAs)!
- 9. ©2016 CodeKaram8 • Document, understand and prioritize SLAs • Throughput • Responsetime • Capacity • Footprint • CPU utilization? • … Define Success!
- 10. ©2016 CodeKaram9 Monitor, measure and define performance in terms of throughput, latency, capacity, footprint, utilization … Work For It!
- 14. Throughput ©2016 CodeKaram13 • Expected throughput? • Can you fall below the expected throughput? • How long can you stay below the expected throughput? • What is the lowest that you can go?
- 15. Throughput ©2016 CodeKaram14 • Throughput measurement • How is it measured? • Transactions/sec; messages/sec or all of them? • Where is it measured? • Client-side; server-side; browser?
- 18. ResponseTime ©2016 CodeKaram17 • Expected response time? • Can you go above the expected response time? • How long can you stay there? • How much can you exceed?
- 19. ResponseTime ©2016 CodeKaram18 • Response time measurement • How is it measured? • 99th percentile; 5-9s; worst-case; or all? • Where is it measured? • Client-side; server-side? Complete loop?
- 23. Capacity Management ©2016 CodeKaram22 • What is the expected load? • What happens if one system gets loaded more than others? (load balancer issue) • How is it measured?
- 24. Capacity Management ©2016 CodeKaram23 • What’s the max load that a system and all systems can tolerate? • How long? • What metrics are being captured?
- 27. Performance Analysis ©2016 CodeKaram26 • Analyze what factors enable the end-user experience to meet or exceed the promised QoS • Track your SLAs!
- 28. Java Application Stack ©2016 CodeKaram27 OS Hardware Application Ecosystem Application JRE JVM
- 29. ©2016 CodeKaram28 • Application services • Application server • Database • Any other services in the ecosystem? Application Performance Analysis Application Ecosystem Application
- 30. JVM + Runtime Performance Analysis ©2016 CodeKaram29 JRE JVM • Classloading stats • JIT Compilation stats • Garbage Collection stats • Threads stats
- 31. OS Performance Analysis ©2016 CodeKaram30 • System/ Kernel stats • Lock stats • Threads stats OS
- 32. Hardware Performance Analysis ©2016 CodeKaram31 • Memory bandwidth/ traffic/ consumption • CPU/ core utilization • CPU cache efficiency/ utilization/ levels • Architectural specific? • IO Stats Hardware
- 33. What areYouTrying to Achieve? ©2016 CodeKaram32 Improve application?
- 34. Top-Down Approach ©2016 CodeKaram33 OS Hardware Application Ecosystem Application JRE JVM
- 35. Top Down Approach - Process •Step 1Monitor •Step 2 Profile •Step 3Analyze •Step 4 Tune + Apply ©2016 CodeKaram34
- 36. What areYouTrying to Achieve? ©2016 CodeKaram35 Improve the platform?
- 37. Bottom-Up Approach ©2016 CodeKaram36 OS Hardware Application Ecosystem Application JRE JVM
- 38. Bottom Up Approach - Process •Step 1Monitor •Step 2 Profile •Step 3Analyze •Step 4 Tune + Apply ©2016 CodeKaram37
- 40. Top-Down Approach ©2016 CodeKaram39 I HAVE the power!! … to modify the code
- 41. Top-Down Approach ©2016 CodeKaram40 OS Hardware Application Ecosystem Application JRE JVM
- 42. Top-Down Approach ©2016 CodeKaram41 Application Ecosystem Application OS Hardware JRE JVM Platform
- 43. Step 1: Monitor ©2016 CodeKaram42 Monitor and plot SUT (System UnderTest) statistics
- 44. ©2016 CodeKaram43 Step 1: Monitor
- 45. ©2016 CodeKaram44 Step 1: Monitor
- 46. ©2016 CodeKaram45 Step 1: Monitor
- 47. ©2016 CodeKaram46 Step 1: Monitor
- 48. ©2016 CodeKaram47 Step 1: Monitor
- 49. ©2016 CodeKaram48 • Tools? • VisualVM, Java Flight Recorder • PrintCompilation, PrintGCDetails (+PrintGCDateStamps), jmap – clstats, jcmd GC.class_stats Step 1: Monitor
- 50. ©2016 CodeKaram49 • Tools? • Linux – mpstat, sysstat – iostat,pidstat…, prstat, vmstat, dash, CPU-Z, cacti … • Windows – Performance Monitor,Task Manager, Resource Monitor, CPU-Z, cacti ... Step 1: Monitor
- 51. Step 2+3: Profile + Analyze ©2016 CodeKaram50 • You have all the data that you need! • Identify areas of improvement • Profile those potential hotspots • Analyze those hotspots
- 52. ©2016 CodeKaram51 • Tools? (Free/Open source/GPL/BSD) • Oracle Solaris Studio PerformanceAnalyzer, perf tools, PAPI,CodeXL, Dtrace, Oprofile, gprof, LTT (linux trace toolkit) • Java Application – VisualVM, Netbeans profiler, jconsole … Step 2+3: Profile + Analyze
- 53. Step 4:Tune ©2016 CodeKaram52 • Tune the JVM/GC – select the right heap, the right GC algorithm • Age objects appropriately • Promote only long-lived objects • GC worker threads perVM (for stop-the-world GC events) • GC concurrent worker threads perVM
- 54. Step 4:Tune ©2016 CodeKaram53 • Tune the JVM/GC – select the right heap, the right GC algorithm • See if compressed oops can be helpful • Larger heaps may need AlwaysPretouch to be enabled and also UseLargePages of appropriate size
- 55. Step 4:Tune ©2016 CodeKaram54 • Tune your code to meet or exceed your SLAs • Appropriate ramp-ups and ramp-downs • Object age and retention strategies • Understand what forms your LDS (live data set) • Confirm you are measuring the right thing! J
- 57. Bottom Up Approach ©2016 CodeKaram56 I NEED the power!! … to stress the platform
- 58. Bottom-Up Approach ©2016 CodeKaram57 OS Hardware Application Ecosystem Application JRE JVM
- 59. Bottom-Up Approach ©2016 CodeKaram58 Application Ecosystem Application OS Hardware JRE JVM Platform
- 60. Where to Start? ©2016 CodeKaram59 • Know what you are stressing • Get/ write the appropriate workload/ application • Get/write the appropriate tools
- 61. Know WhatYou Are Stressing! ©2016 CodeKaram60 CPU – Gather performance counter information for your CPU stats, core stats, cache hits, misses and levels, branch predictions, pipeline information, order-of- execution, load-store unit load and queues,etc
- 62. Know WhatYou Are Stressing! ©2016 CodeKaram61 Memory – Gather performance counter information for memory utilization, memorybandwidth, read-write stats, max read bandwidth, max write bandwidth, max cross traffic bandwidth, architectural related considerations, max capacity, etc
- 63. Know WhatYou Are Stressing! ©2016 CodeKaram62 JVM / GC– Gather information related to the change – e.g. new GC! Gather information on different GC phases, parallel work queues and work performance, concurrent work and pressure,internal queues and buffers, any GC work that’s staged?, etc
- 64. Where to Next? ©2016 CodeKaram63 • Know what you are stressing • Get/ write the appropriate workload/ application • Get/write the appropriate tools
- 65. Let’s have some fun!
- 66. Further Reading