A Paradigm Shift: The Increasing Dominance of Memory-Oriented Solutions for High Performance Data Access
2 likes5,478 views
This document discusses the increasing dominance of memory-oriented solutions for high-performance data access. It notes that database lookups are around 20 milliseconds while hashmap lookups are around 20 nanoseconds. It then discusses how abstraction improves software but hurts performance. It outlines the traditional database architecture with disk storage and compares it to newer in-memory and distributed in-memory architectures that can provide faster performance by avoiding disk I/O and leveraging memory and distribution.
A Paradigm Shift: The Increasing Dominance of Memory-Oriented Solutions for High Performance Data Access
- 1. A Paradigm Shift: The Increasing Dominance of Memory-Oriented Solutions for High Performance Data Access! Ben Stopford : RBS
- 2. How fast is a HashMap lookup? ~20 ns
- 3. That’s how long it takes light to travel a room
- 4. How fast is a database lookup? ~20 ms
- 5. That’s how long it takes light to go to Australia and back
- 6. 3 times
- 7. Computers really are very fast!
- 8. The problem is we’re quite good at writing software that slows them down
- 11. There are many reasons why abstraction is a good idea… …performance just isn’t one of them
- 12. Question: is it fair to compare a Database with a HashMap?
- 13. Not really…
- 14. Key Point On one end of ..on the other sits the scale sits the the database… HashMap… …but it’s a very very long scale that sits between them.
- 19. Traditional Shared Shared In Memory Disk Nothing Distributed Simpler In Memory Contract
- 20. Simplifying the Contract
- 21. How big is the internet? 5 exabytes (which is 5,000 petabytes or 5,000,000 terabytes)
- 22. How big is an average enterprise database 80% < 1TB (in 2009)
- 24. Databases have huge operational overheads Taken from “OLTP Through the Looking Glass, and What We Found There” Harizopoulos et al
- 25. Avoid that overhead with a simpler contract and avoiding IO
- 26. Improving Database Performance ! Shared Disk Architecture Shared Disk
- 27. Improving Database Performance ! Shared Nothing Architecture
- 28. Each machine is responsible for a subset of the records. Each record exists on only one machine.! 1, 2, 3… 97, 98, 99… 765, 769… 169, 170… Client 333, 334… 244, 245…
- 29. Improving Database Performance (3)! In Memory Databases! (single address-space)
- 30. Databases must cache subsets of the data in memory Cache
- 31. Not knowing what you don’t know 90% in Cache Data on Disk
- 32. If you can fit it ALL in memory you know everything!!
- 33. The architecture of an in memory database
- 34. Memory is at least 100x faster than disk ms μs ns ps 1MB Disk/Network 1MB Main Memory 0.000,000,000,000 Cross Continental Main Memory L1 Cache Ref Round Trip Ref Cross Network L2 Cache Ref Round Trip * L1 ref is about 2 clock cycles or 0.7ns. This is the time it takes light to travel 20cm
- 35. Memory allows random access. Disk only works well for sequential reads
- 36. This makes them very fast!!
- 37. The proof is in the stats. TPC-H Benchmarks on a 1TB data set
- 38. So why haven’t in memory databases taken off?
- 39. Address-Spaces are relatively small and of a finite, fixed size
- 40. Durability
- 41. One solution is distribution
- 42. Distributed In Memory (Shared Nothing)
- 43. Again we spread our data but this time only using RAM. 1, 2, 3… 97, 98, 99… 765, 769… 169, 170… Client 333, 334… 244, 245…
- 44. Distribution solves our two problems
- 45. We get massive amounts of parallel processing
- 46. But at the cost of loosing the single address space
- 47. Traditional Shared Shared In Memory Disk Nothing Distributed Simpler In Memory Contract
- 48. There are three key themes here: Simplify the Distribution No Disk contract Improve Gain scalability by scalability picking All data is through a appropriate held in RAM distributed ACID architecture properties.
- 49. ODC
- 50. ODC – Distributed, Shared Nothing, In Memory, Semi-Normalised, Graph DB 450 processes 2TB of RAM Messaging (Topic Based) as a system of record (persistence)
- 51. ODC represents a balance between throughput and latency
- 52. What is Latency?
- 54. Which is best for latency? Shared Nothing (Distributed) Traditional In-Memory Database Database Latency?
- 55. Which is best for throughput? Shared Nothing (Distributed) Traditional In-Memory Database Database Latency?
- 56. So why do we use distributed in memory? Plentiful In Memory hardware Latency Throughput
- 57. This is the technology of the now. So what is the technology of the future?
- 60. New Innovations on the Horizon
- 61. These factors are remolding the hardware landscape to one where memory both vast and durable
- 62. This is changing the way we write software
- 63. Huge servers in the commodity space are driving us towards single process architectures that utilise many cores and large address spaces
- 64. We can attain hundreds of thousands of executions per second from a single process if it is well optimised.
- 65. “All computers wait at the same speed” !
- 66. We need to optimise for our CPU architecture ms μs ns ps 1MB Disk/Network 1MB Main Memory 0.000,000,000,000 Cross Continental Main Memory L1 Cache Ref Round Trip Ref Cross Network L2 Cache Ref Round Trip * L1 ref is about 2 clock cycles or 0.7ns. This is the time it takes light to travel 20cm
- 67. Tools like Vtune allow us to optimise software to truly leverage our hardware
- 68. So what does this all mean?
- 69. Further Reading