![Firebird Scalability Dmitry Yemanov [email_address] The Firebird Project http://www.firebirdsql.org](https://image.slidesharecdn.com/firebirdscalability-091001110107-phpapp02/85/Firebird-Scalability-by-Dmitry-Yemanov-in-English-1-320.jpg)
Firebird Scalability, by Dmitry Yemanov (in English)
- 1. Firebird Scalability Dmitry Yemanov [email_address] The Firebird Project http://www.firebirdsql.org
- 2. Agenda Classic Server vs Super Server comparison, limits, solutions Memory usage page and sorting buffers CPU and concurrency multi-CPU and multi-core, TPC-C figures I/O specifics possible improvements Scaling up vs scaling out clustering possibilities A36 Dmitry Yemanov Slide
- 3. Architectural Differences Classic Server Process per connection, single-threaded Native scheduling Private caches Shared database file usage Shared lock table Super Server Single process, multi-threaded Cooperative scheduling Shared caches Exclusive database file usage In-process lock table A36 Dmitry Yemanov Slide
- 4. Classic Server: Limiting Factors Page cache Balance between available RAM and load Frequent modifications = extra page writes Lock manager bottlenecks Single mutex, LockAcquireSpins on SMP Lock access time, LockHashSlots LockMemSize: 100 * <connections> * <cache pages> Other limits Signal delivery (POSIX) Desktop heap size (Windows) A36 Dmitry Yemanov Slide
- 5. Super Server: Limiting Factors Page cache Total amount, 32-bit is not enough Many pages (old FB versions) Cooperative thread scheduling One request active at a time Rescheduling points Other limits Socket pool limits (1024 on Windows) Stack size (~1000 WNET/XNET connections) A36 Dmitry Yemanov Slide
- 6. A Mixture of Both Key ideas behind Single process, multi-threaded Thread pooling Private caches per connection Communication via the lock manager In-process lock table Implementations Vulcan “No-Shared-Cache” mode RedSoft “SuperClassic” architecture Firebird “Transition” branch A36 Dmitry Yemanov Slide
- 7. Memory Usage Page cache Few pages = many I/O operations Many pages = possible problems in CS Remember about the metadata in CS Efficient implementation in SS Temporary space The bigger the better Should be used with care in CS Other Lock and event tables (shared memory) A36 Dmitry Yemanov Slide
- 8. Page Cache Performance for Writers A36 Dmitry Yemanov Slide
- 9. CPU Scalability A36 Dmitry Yemanov Slide
- 10. Concurrency: TPC-C benchmark A36 Dmitry Yemanov Slide
- 11. I/O Specifics Tablespaces Data separated from indexes and blobs Separated user data (per schema) Partitioning Active vs backup, by warehouse, etc Underlying improvements Asynchronous I/O Multi-page reads Cache prefetch A36 Dmitry Yemanov Slide
- 12. Scaling Up: Searching for the Best Setup Firebird 2.0 Classic on Linux Use a 8K or 16K page size Find (experimentally) an optimal page cache size Calculate and set up LockMemSize accordingly Play with LockAcquireSpins and test the performance Consider FW=OFF with MaxUnflushedWrites = 1 Set up SortMemUpperLimit, if necessary Monitor the lock table and increase LockHashSlots A36 Dmitry Yemanov Slide
- 13. Scaling Out: Firebird Cluster A36 Dmitry Yemanov Slide fb_inet_server fb_inet_server fb_inet_server fb_inet_server Lock Table Database
- 14. Scaling Out: Firebird Cluster A36 Dmitry Yemanov Slide fb_inet_server fb_inet_server fb_inet_server fb_inet_server Distributed Lock Manager e.g. RedHat DLM Database LT LT LT LT
- 15. The Firebird Project www.firebirdsql.org A36 Dmitry Yemanov Slide