High Performance Erlang - Pitfalls and Solutions
2 likes•744 views
The document discusses high-performance programming in Erlang, particularly focusing on the pitfalls and solutions experienced by the MZSpeed team at Machine Zone, Inc. It covers issues related to message passing, performance challenges with ETS (Erlang Term Storage), and highlights several strategies for optimization, such as sharding, message batching, and adopting domain-specific solutions. Additionally, profiling techniques and a case study on counters provide insights into achieving higher performance with Erlang applications.
![Benchmark
Environment
• Platform
• 2 CPU sockets
• CPU - Intel(R) Xeon(R) CPU E5-2697 v3 @ 2.60GHz
• 14 hardware threads per socket
• 2 hyper-threaded threads for each hardware thread
• Erlang Runtime
Erlang/OTP 18 [erts-7.1] [source] [64-bit] [smp:56:56]
[async-threads:10] [kernel-poll:false]
34Machine Zone, Inc](https://image.slidesharecdn.com/9d0d1063-f0ec-4bfd-9c6f-3fc7af040286-160315054652/85/High-Performance-Erlang-Pitfalls-and-Solutions-34-320.jpg)
High Performance Erlang - Pitfalls and Solutions
- 1. High Performance Erlang: Pitfalls and Solutions MZSPEED Team Machine Zone, Inc. 1Machine Zone, Inc
- 2. Agenda • Erlang at Machine Zone • Issues and Roadblocks • Lessons Learned • Profiling Techniques • Case Study: Counters 2Machine Zone, Inc
- 3. Erlang at Machine Zone • High performance data processing system • How to make it fast? 3Machine Zone, Inc
- 4. 4 Typical Approach: Pipelined tasks Execute different logic in different processes receiver processor sender Machine Zone, Inc
- 5. How do we exchange data between processes? How do we make it fast? 5Machine Zone, Inc
- 7. 0 5 M 10 M 15 M 0 2 4 6 8 10 12 14 16 Updateoperations(total) Numberofprocessesupdating Timeline expected processes updating 0 5 M 10 M 15 M 0 2 4 6 8 10 12 14 16 Updateoperations(total) Numberofprocessesupdating Timeline expected processes updating update operations 7Machine Zone, Inc ETS: update_counter, 10 Keys: Expected vs Observed Performance • Observed Performance is far worse than Expected
- 8. 8Machine Zone, Inc ETS: update_counter, Multiple Keys: Observed Performance 0 5 M 10 M 15 M 0 2 4 6 8 10 12 14 16 Updateoperations(total) Numberofprocessesupdating Timeline 1 key 10 keys 100 keys 1000 keys • Performance is degraded even with multiple keys
- 9. 9Machine Zone, Inc ETS has Locks • Fine-grained locks • 64 locks per table by default
- 10. Message Passing 10Machine Zone, Inc
- 11. 1 10 100 1000 N 1 10 100 M 11Machine Zone, Inc Message Passing: M Processes to N Processes 10:100
- 12. 12Machine Zone, Inc Message Passing: Scheduling Locks • Add to runq with locks
- 13. 13Machine Zone, Inc Message Passing: Scheduler Communication Locks • Task stealing with locks
- 14. 14Machine Zone, Inc Message Passing • Limit on #messages delivered per second • < 10M msgs/sec on MacBook Pro • < 22M msgs/sec on a powerful 56-core server • Independent of #processes
- 15. What if we want to achieve more? 15Machine Zone, Inc
- 16. 16Machine Zone, Inc Solution 1: Minimize Message Exchange Do everything in the same process and shard receiver processor sender receiver processor sender receiver processor sender
- 17. 17Machine Zone, Inc Solution 1 Speed • 100x the #ETS updates • 20x–30x the #message exchanges • No decrease with increase in #processes 0 20 M 40 M 60 M 80 M 100 M 0 20 40 60 80 100 Iterationspersec Processes iterations processes
- 18. 18Machine Zone, Inc Solution 2: Message Batching Batch messages: 5x–30x more msgs/sec 1 10 100 1000 N 1 10 100 M 10:100
- 19. 19Machine Zone, Inc Solution 3: Domain-‐Specific Optimizations • Deliver one message to multiple recipients: 1. Receive, process, store to shared memory using NIF 2. Read from the shared memory, send using push notification sender receiver processor shared memory Total throughput of 600M msgs/sec to many recipients shared memory
- 20. 20Machine Zone, Inc High Performance Erlang: Lessons Learned
- 21. 21Machine Zone, Inc Lesson 1: Use Sharding • Avoid having single process • Use a pool of non-communicating shards • erlang:phash2(Key, NumberOfShards) Shard 2 Shard 1 receive receive send send No communication
- 22. 0 2 M 4 M 6 M 8 M 10 M 12 M 0 1 M 2 M 3 M 4 M 5 M Timetaken(us) Number of insertions ets set maps process dictionary 22Machine Zone, Inc Lesson 2: Use Process Dictionary • Quicker insertions in process dictionary compared to maps
- 23. 0 5 k 10 k 15 k 20 k 25 k 1 10 100 1000 10000 100000 Timefor100kupdates(ms) Number of elements R18 maps R18 process dictionary 23Machine Zone, Inc Lesson 2: Use Process Dictionary • Quicker update operations in process dictionary compared to maps (Erlang 18)
- 24. 24Machine Zone, Inc Lesson 3: Use Limited #Processes • Don’t create too many processes • pid2proc lock for process resolution • Scheduler has to manage the processes • Fixed number of schedulers • Don’t spawn/stop processes frequently • Process resolution again
- 26. 26Machine Zone, Inc pstack: Stack Sampling • pstack (gdb): sampling call stack over multiple threads • $ pstack 1234 • $ gdb -p 1234 • (gdb) thread apply all bt 25
- 27. 27Machine Zone, Inc EEP: Erlang Easy Profiling • https://github.com/virtan/eep • Based on dbg module, low overhead trace ports • Easy to use • Kcachegrind visualization Basic Usage • Collect runtime data to local file • > eep:start_file_tracing("abc"), timer:sleep(10000), eep:stop_tracing(). • Convert to callgrind format • > eep:convert_tracing(“abc"). • Visualize with kcachegrind • $ kcachegrind callgrind.out.abc
- 28. 28Machine Zone, Inc EEP: Example Visualization
- 29. 29Machine Zone, Inc oprofile: System Level Profiling • Initialization $ opcontrol –deinit $ echo 0 > /proc/sys/kernel/nmi_watchdog • Profiling $ operf --vmlinux /usr/lib/debug/vmlinux -- callgraph ./rebar eunit apps=perf_test • Reporting and Visualization $ opreport -cgf | ./gprof2dot.py -f oprofile --show-samples | dot -Tpng -o output.png
- 30. 30Machine Zone, Inc oprofile: Example
- 31. Case Study: Counters 31Machine Zone, Inc
- 32. 32Machine Zone, Inc Need for High Performance Counters • Thousands of global counters • Updated by several hundred processes every second
- 33. 33Machine Zone, Inc Libraries Used • Exometer • ETS • Folsom
- 34. Benchmark Environment • Platform • 2 CPU sockets • CPU - Intel(R) Xeon(R) CPU E5-2697 v3 @ 2.60GHz • 14 hardware threads per socket • 2 hyper-threaded threads for each hardware thread • Erlang Runtime Erlang/OTP 18 [erts-7.1] [source] [64-bit] [smp:56:56] [async-threads:10] [kernel-poll:false] 34Machine Zone, Inc
- 35. 0 1 M 2 M 3 M 4 M 5 M 6 M 7 M 10 20 30 40 50 60 Numcounterupdatesperprocess Num processes ets counter exometer counter exometer fast counter folsom counter 35Machine Zone, Inc Benchmark Results • As number of processes increases, number of counter updates per process decreases
- 36. 36Machine Zone, Inc Limitations • Exometer: • counters – scales well but slow • fast counters – doesn’t scale • Lock contention in ets:update_counter()
- 38. 38Machine Zone, Inc Refresher on Cacheline Source: http://www.slideshare.net/yzt/modern-‐cpus-‐and-‐caches-‐a-‐starting-‐point-‐for-‐programmers/52
- 39. 39Machine Zone, Inc Occurs when two CPUs access different data structures on the same cache line False Sharing (SMP) CPUO CPU1 struct foo struct bar Cache line
- 40. 40 CPUO Main memory L2 L2 CPU1 foo bar foo bar foo bar FETCH foo FETCH bar FETCH foo FETCH bar False Sharing (SMP) Machine Zone, Inc
- 41. Main memory 41 False Sharing (SMP) CPUO L2 L2 CPU1 foo bar WRITE foo WRITE foo INVL bar INVL bar foo bar foo bar ACK ACK ACK ACK Machine Zone, Inc
- 42. Main memory 42 CPUO L2 L2 CPU1 WRITE bar WRITE bar INVL foo INVL foo foo bar ACK ACK ACK ACK Machine Zone, Inc False Sharing (SMP)
- 43. 43Machine Zone, Inc Benefits of Per-‐CPU Data Source: http://lwn.net/Articles/256433/ Time taken for 500 million counter increment operations on Intel Core 2 QX 6700
- 44. 44Machine Zone, Inc Benefits of Per-‐CPU Data: Reasons • Less overhead of cache coherency protocol • Significantly higher benefits with QPI (Quick Path Interconnect)
- 45. 45Machine Zone, Inc Transactional Memory • Intel TSX (Transactional Synchronizations Extensions) • Helps with locks in small critical sections • Not yet available in x86 (disabled due to hardware bugs)
- 46. 46Machine Zone, Inc MZMETRICS • Design based on per-CPU counters • Counter Operations: • Create a new counter – requires lock • Read/update counter – no lock required • Delete counter – requires lock • Implemented using NIF
- 47. 47Machine Zone, Inc MZMETRICS Evaluation 0 2 M 4 M 6 M 8 M 10 M 12 M 14 M 16 M 18 M 20 M 10 20 30 40 50 Numcounterupdatesperprocess Num processes exometer mzmetrics • 10X faster than Exometer counters
- 48. 48Machine Zone, Inc MZMETRICS Evaluation: Results • 1 million counter updates with 48 processes • Exometer takes ~ 1 seconds • MZMETRICS takes ~0.1 seconds
- 49. 49Machine Zone, Inc Takeaways • Pitfalls • Message passing becomes expensive at scale • Large number of processes needing pid resolution from a global table causes contention • Existing libraries not amiable for fast counting • Proposed Solutions • Use sharding & Limit messaging passing • Limit number of active processes • Use MZMETRICS for fast counters
- 50. 50Machine Zone, Inc We are Hiring http://www.machinezone.com/careers