IMDB_Scalability
Download as PPTX, PDF0 likes82 views
This document discusses approaches to scaling in-memory databases on multicore hardware. There are two main approaches: employing a symmetric database engine where a single process uses multiple threads across cores to access shared memory, and employing a partitioned database engine where the database is divided into partitions managed by dedicated cores. A challenge is that cache coherency limits scalability as it does not scale to thousands of cores. The document recommends a software-hardware co-design approach, avoiding centralized critical sections, leveraging hardware message passing, and using techniques like optimistic concurrency control to improve scalability on high core count systems.
IMDB_Scalability
- 1. The Goal: In-Memory Database Scalability core core ore core core core core core core core ore core core core core core core core ore core core core core core core core ore core core core core corecore Multicores (CMP) Multisocket Multicores How can DBMS engine scale across hardware islands ? in-memory database on multicore
- 2. Two Approaches to In-Memory Database Scalability on Multicore Employing a Symmetric Database Engine Example Systems: SILO and Hekaton Employing a Partitioned Database Engine Example Systems: DORA, PLP, ATraPos and H-Store in-memory database on multicore
- 3. A Symmetric Database Engine Architecture Symmetric Database Engine using Shared Memory Database Process consists of multiple transaction threads each running on a separate core Transaction threads access the entire shared memory with cache coherency enabled Transaction Engine takes care about Serializability Processor takes care about Cache Coherency in-memory database on multicore
- 4. A Partitioned Database Engine Architecture Partitioned Database as Primary Scalability Mechanism Database is divided into multiple partitions each managed by a dedicated core A partition is owner of its data and indexes A centralized routing table is required to route transactions to partitions Multi-partition transactions is executed by migrating transactions between cores Partitioning may be static or dynamic dynamic re-partitioning mechanism enables to load balance partition access All partitions must reside in shared memory with Cache Coherency enabled in-memory database on multicore
- 5. Problem: Cache Coherency Limits Scalability Cache Coherency Mechanism (CCM) limits processor scalability Cache Coherency of thousand cores is not feasible Problem: How can a DBMS engine scale in the presence of non-coherent hardware islands ? Hardware Island with CCM Hardware Island with CCM CCM in-memory database on multicore
- 6. In-memory Database on Thousand cores The paper: Staring Into The Abyss: An Evaluation Of Concurrency Control With One Thousand Cores concludes with the following observation Software-hardware co-design is the only solution to address the performance issues resulting from high core count. For certain functionalities, specialized hardware can significantly improve performance while reducing power consumption The next slides provides some design guidelines in-memory database on multicore
- 7. System Design Principles Avoid many of the centralized critical sections present in classic database engines (see next slide) Be sensitive to the cost of Cache Coherency (CC) Do you really need CC enabled on all cores? Leveraging hardware message passing for efficient thread synchronization (inter-core communication); example the TILE-Gx processor family in-memory database on multicore
- 8. Avoid Centralized Critical Sections Avoid centralized contention points and make all synchronization scale with the data, allowing larger databases to support more concurrency. Use lightweight locking mechanism and eliminate the bottleneck of a centralized lock manager, by co- locating locks with each DB record Use optimistic based concurrency control possibly combined with multi-version mechanism Use transaction’s local memory - private workspace - to cache its write-set and avoid installing writes until commit time in-memory database on multicore