Advancements in Network and Database Integration

This presentation discusses the convergence of network advancements and database technologies, highlighting key challenges such as packet processing at increasing line rates and latency issues. It explores the current landscape of databases from a research perspective and proposes solutions for optimizing performance in high-speed networking scenarios. The content covers classical network stack models and application-agnostic packet processing architectures to enhance the efficiency of network operations in modern infrastructures.

• Network advancements
• Database integration
• Packet processing
• Latency issues
• Performance optimization

khen Follow

Uploaded on Sep 23, 2024 | 0 Views

Download Presentation

Please find below an Image/Link to download the presentation.

The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author. Download presentation by click this link. If you encounter any issues during the download, it is possible that the publisher has removed the file from their server.

Presentation Transcript

1. Its Time to Combine Network It s Time to Combine Network Advances and Databases Advances and Databases Helge Reelfs CoNEXT 2016, Irvine, CA, US, December 2016 http://comsys.rwth-aachen.de/ Picture from diana colors @ flickr.com

2. Databases Databases Where are we now? Where are we now? Databases exist since 70 s Databases exist since 70 s Still very active research branch Still very active research branch Distribution (scaling out) Distribution (scaling out) Hardware Hardware Software Parallelization Parallelization Software 2 Jens Helge Reelfs

3. But what about networking? But what about networking? Increasing line rates challenge packet processing Increasing line rates challenge packet processing CPU speeds do not scale with increasing line rates Performance problems at high line rates (e.g. >>10Gbps) Main overhead factors Main overhead factors Memory allocations and copy operations System calls and context switches Does it matter after all? Socket API Transmission User-Kernel Copy Protocol Stack Driver Reception NIC 0 1000 2000 3000 4000 5000 [ns] Source: Larsen et al., Architectural Breakdown of End-to-End Latency in a TCP/IP network, J Parallel Prog, 37:6, (2009) 3 Jens Helge Reelfs

4. The current answer is The current answer is R Remote emote D Direct irect M Memory emory A Access ccess Local Database RDMA Cluster Local Database RDMA Cluster TCP/IP/Ethernet Clients & Cloud TCP/IP/Ethernet Clients & Cloud Caching Pool Webserver Pool Database Database Webserver Memcached Kernel Kernel Webserver Memcached NIC NIC NIC NIC Webserver Memcached RDMA TCP/IP TCP/IP NIC NIC other NIC NIC Kernel Kernel other Database Database other Other Clients 4 Jens Helge Reelfs

5. Classical Network Stack (simplified) Classical Network Stack (simplified) Result Result VALUE <key> <value> END Web- Mem- cached server MySQL remove remove it! it! user CTX SW Copy kernel TRANS NET kernel CTX SW Copy hardware MAC Request Request GET <key> END 5 Jens Helge Reelfs

6. Application Agnostic Packet Processor Application Agnostic Packet Processor - - Santa Santa Z ZZ ZZ Z Mem- cached control plane user kernel Result Result VALUE <key> <value> END TRANS Santa NET kernel CTX SW Copy hardware MAC Request Request GET <key> END 6 Jens Helge Reelfs

7. Preliminary Tests Preliminary Tests 4 (Memcached Memcached) or 1 (MySQL MySQL) clients Uniform workload (memaslap & mysqlslap) Server + Santa Kernel 10 10 Gbps Modified Modified Applications Applications using UDP UDP Transport: Memcached TCP TCP Transport: MySQL Gbps links links using Santa Santa kernel kernel API API 7 Jens Helge Reelfs

8. UDP Transport: UDP Transport: Memcached Memcached Results Results * Motivated Motivated by Santa reduces latency, increases throughput Santa reduces latency, increases throughput Leveling equally @ line rate (1 KB payload) Leveling equally @ line rate (1 KB payload) Leveling equally @ line rate (1 KB payload) Leveling equally @ line rate (1 KB payload) Motivated by real Motivated by real- -world: very small values by real real- -world world: : very world: very small values very small small values values Santa reduces latency, increases throughput Santa reduces latency, increases throughput * Scaling Memcached at Facebook @ NSDI 13 8 Jens Helge Reelfs facebook logo @ facebook.com

9. TCP Transport: MySQL Results TCP Transport: MySQL Results MySQL Query Cache is very efficient MySQL Query Cache is very efficient Santa further improves throughput Santa further improves throughput Leveling equally @ higher payloads Leveling equally @ higher payloads Leveling equally @ higher payloads Leveling equally @ higher payloads MySQL Query Cache is very efficient MySQL Query Cache is very efficient Santa further improves throughput Santa further improves throughput 9 Jens Helge Reelfs

10. Future Work Future Work Improve Santa Improve Santa Use eBPF Push processor down towards hardware Evaluation of different applications Evaluation of different applications In-Memory Database Distributed Database Evaluation of Kernel Evaluation of Kernel- -Bypassing implementations Bypassing implementations StackMap Evaluation of different workloads Evaluation of different workloads Power-law Transactional Evaluation of latency impact Evaluation of latency impact 10 Jens Helge Reelfs

11. Thanks. Thanks. Helge Reelfs CoNEXT 2016, Irvine, CA, US, December 2016 http://comsys.rwth-aachen.de/

12. The answer is The answer is R Remote emote D Direct irect M Memory emory A Access ccess Widely used, various implementations (HPC, Widely used, various implementations (HPC, big Concept Concept: lightweight direct data transfer & placement : lightweight direct data transfer & placement Simple primitives (read, write, Simple primitives (read, write, ack Pre Pre- -allocated buffers per connection, polling allocated buffers per connection, polling Astonishing results (latency, throughput, CPU load) Astonishing results (latency, throughput, CPU load) big databases) databases) ack) ) But But Specialized software Specialized hardware (NIC) Specialized networking infrastructure TCP/IP/Ethernet compatible implementations immature* *Zhu et al. Congestion Control for Large-Scale RDMA Deployments (SIGCOMM 15) 12 Jens Helge Reelfs

13. The good News: TCP/IP Alternatives exist! The good News: TCP/IP Alternatives exist! Bypassing Kernel APP Classical Network Stack Kernel- Software APP STACK user kernel netmap / dpdk APP RDMA TRANS TRANS TRANS NET NET NET kernel HW MAC MAC MAC 13 Jens Helge Reelfs