Critical Attributes for a High-Performance, Low-Latency Database
•
0 likes•941 views
This document discusses the attributes of a high-performance, low-latency database like ScyllaDB. It begins with introductions and an overview of ScyllaDB. It then summarizes how hardware has evolved over 20 years with more cores, memory, and faster disks. ScyllaDB was redesigned from first principles to take advantage of modern hardware, using an asynchronous, shared-nothing architecture with one shard per core. This allows it to achieve significantly higher performance than Cassandra. The document shows benchmark results demonstrating ScyllaDB's lower latencies and ability to scale to higher throughput. It also discusses how ScyllaDB uses workload prioritization to manage different types of workloads.
Recommended
Recommended
More Related Content
What's hot
What's hot (20)
Similar to Critical Attributes for a High-Performance, Low-Latency Database
Similar to Critical Attributes for a High-Performance, Low-Latency Database (20)
More from ScyllaDB
More from ScyllaDB (20)
Recently uploaded
Recently uploaded (20)
Critical Attributes for a High-Performance, Low-Latency Database
- 1. Benny Halevy, Director Software Engineering Tzach Livyatan, VP product, ScyllaDB Attributes of a High-Performance, Low-Latency Database
- 2. 2 ~$ whoami Benny Halevy, Director Software Engineering Tzach Livyatan, VP product, ScyllaDB
- 3. 3 Agenda + About ScyllaDB + 20 years of hardware evolution in 5 minutes + Scylla - Design for performance + Results + Workload Prioritization + Summary
- 4. 4 + The Real-Time Big Data Database + Fully Compatible with Apache Cassandra and Amazon DynamoDB + 10X the performance & low tail latency + Open Source, Enterprise and Cloud options + Founded by the creators of KVM hypervisor + HQs: Palo Alto, CA, USA; Herzelia, Israel; Warsaw, Poland About ScyllaDB
- 5. Cluster - Node Ring 5 Node 5 Node 2 Node 1 Node 3 Node 4
- 6. Active/active, replicated, auto-sharded 6 Scylla Architecture
- 7. Why Scylla? On-Prem Cloud Hosted Scylla Cloud Best High Availability in the industry Best Disaster Recovery in the industry Best Scalability in the industry Best Performance in the industry Auto-tune — out of the box performance Fully compatible with Cassandra & DynamoDB The power of Cassandra at the speed of Redis and more
- 8. 8 20 years of hardware evolution in 5 minutes
- 9. Basic architecture - Cassandra 9 Disk MMaped file Kernel tasks Main threadpool M:N threads:clients Client Client Client Page fault
- 10. 10
- 11. 11
- 12. Non Uniform Memory Access (NUMA) 12
- 13. What happened? 13 + Per thread performance plateaued + Cores: 1 ⟶ 256, NUMA + RAM: 2GB ⟶ 2TB + Disk space: 10GB ⟶ 10TB + Disk seek time: 10-20ms ⟶ 20µs + Network throughput: 1Gbps ⟶ 100Gbps This year: 64/128 cores/threads/cpu, 400Gbps NIC, Disk 10µs latency, 1.5TB/device, DDR5 2TB/DIMM AWS u-24tb1.metal: 224 cores, 448 threads, 24TB RAM
- 14. 14 Audience Poll NoSQL Database Adoption
- 15. 15 The database, reimagined Redesigning from first principles
- 16. Shard per core Share nothing, block nothing 16
- 17. Sharding/partitioning + Common concept in distributed databases + Break the system to N non-interacting parts + Usually done by hash(partition_key) % N + Data/load may be unbalanced + Fact of life in distributed databases 🤷 + Logical mapping of data shards to core shards 17
- 18. Sharding all the way down 18 Node ID Shard ID
- 19. Seastar + Open source framework, powering Scylla, Ceph, Redpanda, ValuStor and more + A “mini operating system in userspace” + Task scheduler, I/O scheduler + Fully asynchronous - userspace coroutines + Direct I/O, (bypasses kernel pagecache) + App should implement caching on its own. + One thread per core, one shard per core 19
- 20. Shard per Core Cassandra TCP/IP Scheduler queue queue queue queue queue Threads NIC Queues Kernel Traditional Stack SeaStar’s Sharded Stack Memory Lock contention Cache contention NUMA unfriendly TCP/IP Task Scheduler queue queue queue queue queue smp queue NIC Queue DPDK Kernel (isn’t involved) Userspace TCP/IP Task Scheduler queue queue queue queue queue smp queue NIC Queue DPDK Kernel (isn’t involved) Userspace TCP/IP queue queue queue queue queue smp queue NIC Queue Kernel (isn’t involved) Userspace No contention (*) Linear scaling NUMA friendly (*) cooperative- preemption model in shard Core Database Task Scheduler queue queue queue queue smp queue Userspace NIC Queue 20
- 21. Unified Cache Cassandra Key cache Row cache Linux page cache SSTables App thread Kernel SSD Page fault Suspend thread Initiate I/O Context switch I/O completes Interrupt Context switch Map page Resume thread Page fault On-heap / Off-heap 21 Shared memory; NUMA unfriendly
- 22. Unified Cache Cassandra Scylla Key cache Row cache Linux page cache SSTables Unified cache SSTables Complex Tuning On-heap / Off-heap 22 Async, direct I/O Keys / Rows GP Buffers
- 23. Thou shalt not block Query Commitlog Compaction Queue Queue Queue Userspace I/O Scheduler Disk Max useful disk concurrency I/O queued in FS/device No queues 23
- 24. Memtable Seastar Scheduler Compaction Query Repair Commitlog SSD Compaction Backlog Monitor Memory Monitor Adjust priority NET CPU How does scheduling work? 24
- 25. 25 Shard aware I/O scheduler(s) + Each shard has independent scheduler + Capacity groups per NUMA zone + Shards grab capacity leases Minimal, low cost coordination between shards!
- 27. The controllers - memtable 27 This is the CPU percentage needed (50 %) To keep the buffers at a stable level Throughput barely oscillates Total system CPU usage barely oscillates
- 28. The controllers - memtable 28 without controller with controller
- 31. 31 Write Latency - Scylla vs Cassandra
- 32. 32 Read Latency - Scylla vs Cassandra
- 33. 33 Write Latency - 4 Scylla nodes vs. 40 Cassandra nodes
- 35. 35 Real world results C* nodes 962 Scylla nodes 78 + 5x-10x throughput compared to Cassandra + Vertical scaling to hundreds of CPUs
- 38. Workload Prioritization: Different types of loads ■ OLTP ● Small work items ● Latency sensitive ● involves narrow portion of the data ■ OLAP ● Large work items ● Throughput oriented ● Performed on large amounts of data
- 39. + Shares are really all there is to it :) + Schedulers maintain fairness by trying to optimize ratios and not absolute throughput. + Schedulers only kick in when there is a conflict on the resource. + Schedulers can be dynamic - meaning you can change the amount of shares in real time. + Limits the impact of one Share-Holder on another. Schedulers Basics - operation highlight
- 40. Memtable Seastar Scheduler Compaction Query Repair Commitlog Compaction Backlog Monitor Memory Monitor Adjust priority NET CPU How does it work? SSD 40
- 41. How does it work? 41 Memtable Seastar Scheduler Compaction Query Repair Commitlog Compaction Backlog Monitor Memory Monitor Adjust priority NET CPU SSD
- 42. How does it work? Workload Prioritization! Service-level Controller 42 Memtable Seastar Scheduler Compaction Query Repair Commitlog Compaction Backlog Monitor Memory Monitor Adjust priority NET CPU SSD
- 43. How does it work?
- 45. Configuring Workload prioritization 1. Make users that generates the same workload be part of the same group. ● Priorities are attached to groups or individual users. 2. Create a service level for the workload and set its shares: ● Share determine the amount of importance of the service level. ● It is always relative to other service levels. 3. Attach the service level to the group of users. ● This will grant the shares to the group of users. ● At that point the workload prioritization mechanizm will start to ● Treat their requests according to priorities.
- 46. Managing Workload Prioritization using CQL 1. Make users that generates the same workload be part of the same group. ● CREATE ROLE super_high_priority; ● GRANT super_high_priority TO special_user; 2. Create a service level for the workload and set its shares: ● CREATE SERVICE_LEVEL 'important_load' WITH SHARES=1000; 3. Attach the service level to the group of users. ● ATTACH SERVICE_LEVEL 'important_load' TO ‘super_high_priority;
- 47. Workload Prioritization to the Rescue! ■ Load1: 200 shares, Load2: 400 shares, Load3: 800 shares Shares determine workload latency 47
- 48. 48 + Design and built to meet modern hardware + Use a fully async, share nothing, shard per core architecture + Superior throughput and consistent low latency + Expose internal scheduler to the user as Workload Prioritization Summary
- 50. United States 2445 Faber St, Suite #200 Palo Alto, CA USA 94303 Israel Maskit 4 Herzliya, Israel 4673304 www.scylladb.com @scylladb Thank You!