With the rapid development of NIC, the poor performance of data packets processing with Linux kernel has become the bottleneck. However, the rapid development of the Internet needs high performance of network processing, kernel bypass has caught more and more attention. There are various similar technologies appear, such as DPDK, NETMAP and PF_RING. The main idea of kernel bypass is that Linux is only used to deal with control flow, all data streams are processed in user space. Therefore, kernel bypass can avoid performance bottlenecks caused by kernel packet copy, thread scheduling, system calls and interrupt. Furthermore, kernel bypass can achieve higher performance with multi optimizing methods. Within various techniques, DPDK has been widely used because of its more thorough isolation from kernel scheduling and active community support.
F-Stack is an open source network framework with high performance based on DPDK. With follow characteristics
- Ultra high network performance which can achieve network card under full load, 10 million concurrent connection, 5 million RPS, 1 million CPS.
- Transplant FreeBSD 11.01 user space stack, provides a complete stack function, cut a great amount of irrelevant features. Therefore greatly enhance the performance.
- Support Nginx, Redis and other mature applications, service can easily use F-Stack
- With Multi-process architecture, easy to extend
- Provide micro thread interface. Various applications with stateful app can easily use F-Stack to get high performance without processing complex asynchronous logic.
- Provide Epoll/Kqueue interface that allow many kinds of applications easily use F-Stack
In order to deal with the increasingly severe DDoS attacks, authorized DNS server of Tencent Cloud DNSPod switched from Gigabit Ethernet to 10-Gigabit at the end of 2012. We faced several options, one is to continue to use the original model another is to use kernel bypass technology. After several rounds of investigation, we finally chose to develop our next generation of DNS server based on DPDK. The reason is DPDK provides ultra-high performance and can be seamlessly extended to 40G, or even 100G NIC in the future.
After several months of development and testing, DKDNS, high-performance DNS server based on DPDK officially released in October 2013. It's capable of achieving up to 11 million QPS with a single 10GE port and 18.2 million QPS with two 10GE ports. And then we developed a user-space TCP/IP stack called F-Stack that can process 0.6 million RPS with a single 10GE port.
With the fast growth of Tencent Cloud, more and more services need higher network access performance. Meanwhile, F-Stack was continuous improving driven by the business growth, and ultimately developed into a general network access framework. But this TCP/IP stack couldn't meet the needs of these services while continue to develop and maintain a complete network stack will cost high, we've tried several plans and finally determined to port FreeBSD(11.0 stable) TCP/IP stack into F-Stack. Thus, we can reduce the cost of maintenance and follow up the improvement from community quickly.Thanks to libplebnet and libuinet, this work becomes a lot easier.
With the rapid development of all kinds of application, in order to help different APPs quick and easily use F-Stack, F-Stack has integrated Nginx, Redis and other commonly used APPs, and a micro thread framework, and provides a standard Epoll/Kqueue interface.
Currently, besides authorized DNS server of DNSPod, there are various products in Tencent Cloud has used the F-Stack, such as HttpDNS (D+), COS access module, CDN access module, etc..
#clone F-Stack
mkdir /data/f-stack
git clone https://github.com/F-Stack/f-stack.git /data/f-stack
cd f-stack
# compile DPDK
cd dpdk/tools
./dpdk-setup.sh # compile with x86_64-native-linuxapp-gcc
# Set hugepage
# single-node system
echo 1024 > /sys/kernel/mm/hugepages/hugepages-2048kB/nr_hugepages
# or NUMA
echo 1024 > /sys/devices/system/node/node0/hugepages/hugepages-2048kB/nr_hugepages
echo 1024 > /sys/devices/system/node/node1/hugepages/hugepages-2048kB/nr_hugepages
# Using Hugepage with the DPDK
mkdir /mnt/huge
mount -t hugetlbfs nodev /mnt/huge
# close ASLR; it is necessary in multiple porcess
echo 0 > /proc/sys/kernel/randomize_va_space
# offload NIC
modprobe uio
insmod /data/f-stack/dpdk/x86_64-native-linuxapp-gcc/kmod/igb_uio.ko
insmod /data/f-stack/dpdk/x86_64-native-linuxapp-gcc/kmod/rte_kni.ko
python dpdk-devbind.py --status
ifconfig eth0 down
python dpdk-devbind.py --bind=igb_uio eth0 # assuming that use 10GE NIC and eth0
# Compile F-Stack
cd ../../lib/
make
export FF_PATH=/data/f-stack
export FF_DPDK=/data/f-stack/dpdk/x86_64-native-linuxapp-gcc
cd app/nginx-1.11.10
./configure --prefix=/usr/local/nginx_fstack --with-ff_module
make
make install
cd ../..
./start.sh -b /usr/local/nginx_fstack/sbin/nginx -c config.ini
cd app/redis-3.2.8/
make
make install
Test environment
NIC:Intel Corporation Ethernet Controller XL710 for 40GbE QSFP+
CPU:Intel(R) Xeon(R) CPU E5-2670 v3 @ 2.30GHz
Memory：128G
OS:CentOS Linux release 7.2 (Final)
Kernel：3.10.104-1-tlinux2-0041.tl2
Nginx uses linux kernel's default config, all soft interrupts are working in the first CPU core.
Nginx si means modify the smp_affinity of every IRQ, so that the decision to service an interrupt with a particular CPU is made at the hardware level, with no intervention from the kernel.
CPS (Connection:close, Small data packet) test result
RPS (Connection:Keep-Alive, Small data packet) test data
Bandwidth (Connection:Keep-Alive, 3.7k bytes data packet) test data
See LICENSE