Making the Web Faster with User Level Networking: 2X NGINX Performance Increase for Video on Demand, Load Balancing, Web Content
Download as pptx, pdf0 likes606 views
This presentation explains how three key features of Solarflare's OpenOnload application acceleration stack contribute to increasing Nginx application performance. First, OpenOnload allows network processing at user-level, bypassing the OS kernel. Because OpenOnload includes a TCP stack that runs completely in user space, it avoids the overhead of system calls that consume CPU resources. In addition, application and network performance is improved without sacrificing the security and multiplexing functions that the OS kernel normally provides. Second, OpenOnload now supports the SO_REUSEPORT socket option that allows multiple TCP listening sockets to bind to the same IP address and port and to distribute incoming connection requests between these sockets. This feature can be exploited by applications to increase their scalability. OpenOnload’s implementation of SO_REUSEPORT is superior to that of the Linux kernel because OpenOnload has the ability to instantiate independent, per process, per thread network stacks each of which has direct access to hardware. This eliminates lock contention between processes when sending to and receiving from the network stack. Third, OpenOnload supports socket caching, which avoids the overhead of additional system calls. This improves performance when connections are short-lived.
![[OpenStack Days Korea 2016] How open HW and SW drives telco infrastucture inn...](https://cdn.slidesharecdn.com/ss_thumbnails/03skt-160226163436-thumbnail.jpg?width=560&fit=bounds)
![[OpenStack 하반기 스터디] HA using DVR](https://cdn.slidesharecdn.com/ss_thumbnails/openstackteam2dvr20151214-151221115503-thumbnail.jpg?width=560&fit=bounds)
![[OpenStack Days Korea 2016] Track1 - Red Hat enterprise Linux OpenStack Platform](https://cdn.slidesharecdn.com/ss_thumbnails/13redhat-160226171603-thumbnail.jpg?width=560&fit=bounds)
![[OpenStack Days Korea 2016] An SDN Pioneer's Vision of Networking](https://cdn.slidesharecdn.com/ss_thumbnails/06vmware-160226170659-thumbnail.jpg?width=560&fit=bounds)
![[OpenStack Day in Korea 2015] Track 3-2 - Huawei Cloud Computing Powered by O...](https://cdn.slidesharecdn.com/ss_thumbnails/32-150213065750-conversion-gate02-thumbnail.jpg?width=560&fit=bounds)
![[OpenStack Day in Korea 2015] Track 3-4 - Software Defined Storage (SDS) and ...](https://cdn.slidesharecdn.com/ss_thumbnails/34-150213070106-conversion-gate02-thumbnail.jpg?width=560&fit=bounds)
![[OpenStack Day in Korea 2015] Track 2-3 - 오픈스택 클라우드에 최적화된 네트워크 가상화 '누아지(Nuage)'](https://cdn.slidesharecdn.com/ss_thumbnails/23-150213052422-conversion-gate01-thumbnail.jpg?width=560&fit=bounds)
![[OpenStack Day in Korea 2015] Track 3-6 - Archiectural Overview of the Open S...](https://cdn.slidesharecdn.com/ss_thumbnails/36-150213070342-conversion-gate02-thumbnail.jpg?width=560&fit=bounds)
![[OpenStack Day in Korea 2015] Keynote 5 - The evolution of OpenStack Networking](https://cdn.slidesharecdn.com/ss_thumbnails/05-150213034012-conversion-gate02-thumbnail.jpg?width=560&fit=bounds)
![[OpenStack Day in Korea] Keynote#2 - Bringing OpenStack to the Enterprise Dat...](https://cdn.slidesharecdn.com/ss_thumbnails/koreaopenstackdubuque20140218r7-140218091223-phpapp02-thumbnail.jpg?width=560&fit=bounds)
More Related Content
What's hot (20)
Viewers also liked (10)
Similar to Making the Web Faster with User Level Networking: 2X NGINX Performance Increase for Video on Demand, Load Balancing, Web Content (20)
Recently uploaded (20)
Making the Web Faster with User Level Networking: 2X NGINX Performance Increase for Video on Demand, Load Balancing, Web Content
- 1. January 5th 2015 | 1 Copyright © 2015 Solarflare Communications, Inc. All rights reserved. Making the Web Faster with User Level Networking: 2X Performance Increase for Video on Demand, Load Balancing, Web Content Bruce Tolley, PhD 23 Sept 2015 Nginx Conference San Francisco, CA
- 2. January 5th 2015 | 2 Introducing Solarflare - Who We Are • High-performance 10 and 40GbE software and server adapters • Power the networks at all the banks and exchanges • Partner with NGINX, Arista, Cisco, Red Hat, Juniper • OEM with Dell, HP, Lenovo. Partner with Ciara, Quanta, SuperMicro • Over 1400 customers worldwide • Deployed at CloudFlare, Digital Ocean, LeaseWeb and more • Trials and POCs @ in US, Europe, China
- 3. January 5th 2015 | 3 What We Do We optimize performance from the application down to the wire
- 4. January 5th 2015 | 4 How We Do It--User Level Networking: OpenOnload • Accelerates BSD sockets • TCP/IP, UDP and multicast • Bypasses the kernel stack • Increases throughput • Integrates with existing infrastructure • Binary compatible with standard APIs • TCP/IP and UDP on the wire • Open Source. Bare metal, VM, container • Added support for enhanced socket caching. SO_REUSEPORT
- 5. January 5th 2015 | 5 Solarflare Tools, Libraries, and Interfaces for Dev Teams • EF-VI EtherFabric Virtual Interface • Bare, Layer 2 interface for applications in user space. Similar to DPDK • Future: DPDK • SolarCapture FDK (firmware development kit) • Standards-based interfaces (libpcap) and libraries for home- brewed capture and security applications • OpenOnload • High-performance, Open-Source, Linux-friendly stack that dramatically reduces latency and CPU utilization • IETF RFC compliance. Hardened. Running in 10s of 1000s of nodes: NYSE, NASDAQ
- 6. January 5th 2015 | 6 Onload Architecture and Two New Features to Support Nginx Acceleration • Network processing at user-level, bypassing the OS kernel • Frees up expensive CPU resources • New support of the SO_REUSEPORT socket option that allows multiple TCP listening sockets to bind to the same IP address and port • increases application scalability • Higher SO_REUSEPORT performance than the Linux stack because Onload eliminates lock contention via load balancing • Socket caching, avoids the overhead of additional system calls • improves performance when connections are short-lived
- 7. January 5th 2015 | 7 NGINX Performance: Short Lived Connections- Web Traffic – 10GbE 0 20000 40000 60000 80000 100000 120000 1 2 3 4 5 6 8 10 12 10 Gbps Connections (x axis cores, y axis connection/s) sfn7002f-Onload sfn7002f-kernel x710-kernel Solarflare SFN7002 3X More Efficient than Intel X710xx Line Rate @ 9 Gbps Max Performance Only Limited by the Bandwidth of the 10GbE Link
- 8. January 5th 2015 | 8 NGINX VOD Performance-- Long Lived Connections– 40GbE and 2X10GbE 0 10,000 20,000 30,000 40,000 1 2 3 4 5 6 7 8 Connnections(max.) Cores Nginx Performance 40GbE SFN7002 Onload Intel XL710 5,000 10,000 15,000 20,000 1 2 3 4 5 6 Connections(Max) Cores Nginx Performance 2X10GbE SFN7002 Onload Intel X710 Solarflare: max performance with just two cores. 18K simultaneous connections. Intel: needs three cores to fill up the link Solarflare: maximum performance with 3 cores. Intel: 4 to 5 cores Solarflare 1.67 times more efficient than Intel
- 9. January 5th 2015 | 9 What Other Efficiencies Can Be Found • Power measured using the Dell iDRAC on Dell R610 Server • Idle power same: Intel x710 vs Solarflare SFN7002 @ 70w • 2X more power efficient than Intel
- 10. January 5th 2015 | 10 NGINX Load Balancer Performance Improvements 0 10 20 30 40 50 1 2 3 4 5 6 7 Connections/Se c (000s) Cores 100K Byte i40e SFN7142q 0 50 100 150 200 250 300 350 400 1 2 3 4 5 6 7 8 Connections/Se c (000s) Cores 1K Byte i40e SFN7142q 0 50 100 150 200 250 300 1 2 3 4 5 6 7 Connections/Sec (000s) Cores 10K Byte i40e SFN7142q
- 11. January 5th 2015 | 11 Summary NGINX Performance with Solarflare • Bottom line: Solarflare delivers NGINX acceleration and significant reductions in CPU utilization • For Web traffic (short-lived HTTP connections) • For Video on Demand traffic (long-lived HTTP connections) • Also Nginx Load Balancer • 3 to 4X higher performance than Intel • Power Efficiency • 2X More power efficient than Intel
- 12. January 5th 2015 | 12 Where to Get More Information • Introduction to OpenOnload • https://support.solarflare.com • Solarflare and CloudFlare Build a Faster, Safer, More Secure Web • http://www.solarflare.com/Content/UserFiles/Documents/Solarflare_CloudFlare_CaseStudy.pdf • Improving Web Server Performance • http://www.solarflare.com/Accelerating-Web-CDN-Applications • Accelerating the VOD Work Flow • http://www.solarflare.com/Content/UserFiles/Documents/RTL_HD_Live_to_VOD_and_Playout_ Automated_Workflow.pdf • Solarflare+RTL Webinar • https://www.brighttalk.com/webcast/11615/141217 • “How to achieve Low Latency with 10Gbps” • https://blog.cloudflare.com/author/marek-majkowski/
- 13. January 5th 2015 | 13 END
Editor's Notes
- #5: Accelerates BSD sockets TCP/IP, UDP and multicast acceleration Bypasses the kernel stack Increases throughput Seamlessly integrates into existing infrastructure Binary compatible with industry standard APIs No software modifications are needed Standards-based solution uses TCP/IP and UDP No specialized protocols needed Compatible with existing Ethernet infrastructure All major Linux variants supported Open source, available with bundled support (EnterpriseOnload)
- #7: Three key features of OpenOnload contribute to increasing Nginx application performance. First, OpenOnload allows network processing at user-level, bypassing the OS kernel. Because OpenOnload includes a TCP stack that runs completely in user space, it avoids the overhead of system calls that consume CPU resources. In addition, application and network performance is improved without sacrificing the security and multiplexing functions that the OS kernel normally provides. Second, OpenOnload now supports the SO_REUSEPORT socket option that allows multiple TCP listening sockets to bind to the same IP address and port and to distribute incoming connection requests between these sockets. This feature can be exploited by applications to increase their scalability. OpenOnload’s implementation of SO_REUSEPORT is superior to that of the Linux kernel because OpenOnload has the ability to instantiate independent, per process, per thread network stacks each of which has direct access to hardware. This eliminates lock contention between processes when sending to and receiving from the network stack. Third, OpenOnload supports socket caching, which avoids the overhead of additional system calls. This improves performance when connections are short-lived.
- #8: Solarflare with OpenOnload 3X more efficient than Intel, achieving 10Gbps line rate with only 2 cores, Intel needs 6. Solarflare delivers 2X higher transaction rate than the Intel X710. Figure 1 plots Nginx performance on the y-axis versus the allocated number of cores on the x-axis. Performance is represented as a request rate, in this case, connections per second. The figure shows that as the number of cores is increased (i.e. additional Nginx instances are run) the request rate increases until reaching a ceiling of 111.6 thousand connections per second. The results show that the Solarflare SFN7002 with OpenOnload delivers a maximum 120% increase in Nginx performance over the Intel X710 with Flow Director. With Flow Director off, OpenOnload yields a 128% boost over the Intel X710. When we look at how many CPU cores it takes to saturate a 10 GbE link, OpenOnload needs two cores. In contrast, the Intel X710 adapter (and the Solarflare adapter with its kernel driver) need six or more cores. OpenOnload is therefore found to be three times more efficient than the kernel in its use of CPU resources.