TCP/IP stack for dpdk
ANS(accelerated network stack) is DPDK native TCP/IP stack and also refer to FreeBSD implementation. ANS provides a userspace TCP/IP stack for use with the Intel DPDK.
ans: accelerated network stack process.
librte_ans: TCP/IP stack static library. ANS use dpdk mbuf, ring, memzone, mempool, timer, spinlock. so zero copy mbuf between dpdk and ANS.
librte_anssock: ANS socket lib for application, zero copy between ANS and application.
librte_anscli: ANS cli lib for route/ip/neigh/link configuration.
cli: Command for configure ANS tcp/ip stack.
example: ANS application example.
test: Example application with ANS for testing ANS tcp/ip stack
Support environment
Support feature:
ANS User Guide
# git clone https://github.com/ansyun/dpdk-ans.git
# export RTE_ANS=/home/mytest/dpdk-ans
# ./install_deps.sh
# cd ans
# make
ANS Architecture
TCP Deployment
|-------| |-------| |-------|
APP |anssock| |anssock| |anssock|
|-------| |-------| |-------|
| | |
fd fd fd
--------------------------------------------------
ANS | | |
|-------| |-------| |-------|
| TCP | | TCP | | TCP |
|---------------------------------------|
| IP/ARP/ICMP |
|---------------------------------------|
|LCORE0 | |LCORE1 | |LCORE2 |
|-------| |-------| |-------|
| | |
|---------------------------------------|
| NIC + RSS |
|---------------------------------------|
Performance Testing
|------------------------------|
| TCP Server CPS |
|------------------------------|
| ANS with epoll |
| (one core) |
|------------------------------|
| 100k connection/s |
|------------------------------|
L3 forwarding with NIC performance testing
ENV: CPU- Intel(R) Xeon(R) CPU E5-2683 v3 @ 2.00GHz, NIC- Intel 82599ES 10-Gigabit, Test tool:pktgen-DPDK
|--------------------------------------|
| L3 forwarding performance |
| (one lcore) |
|--------------------------------------|
| Packet size(byte)| Throughput(Mpps) |
|--------------------------------------|
| 64 | 11.78 |
|--------------------------------------|
| 128 | Line Rate |
|--------------------------------------|
L3 forwarding with vhost/virtio performance testing
ENV: Intel(R) Xeon(R) CPU E5-2618L v4 @ 2.20GHz, NIC- vhost/virtio, Test tool:pktgen-DPDK
|--------------------------------------|
| L3 forwarding performance |
| (one lcore) |
|--------------------------------------|
| Packet size(byte)| Throughput(Mpps) |
|--------------------------------------|
| 64 | 6.33 |
|--------------------------------------|
| 128 | 5.94 |
|--------------------------------------|
| 256 | Line Rate |
|--------------------------------------|
====ENV===
CPU:Intel(R) Xeon(R) CPU E5-2609 v4 @ 1.70GHz.
NIC:Ethernet controller: Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection (rev 01)
ANS run on a lcore.
root@ubuntu:~/src/dpdk-redis# ./src/redis-benchmark -h 10.0.0.2 -p 6379 -n 100000 -c 50 -q
PING_INLINE: 138888.89 requests per second
PING_BULK: 141242.94 requests per second
SET: 140449.44 requests per second
GET: 141043.72 requests per second
INCR: 141442.72 requests per second
LPUSH: 141043.72 requests per second
LPOP: 140449.44 requests per second
SADD: 141643.06 requests per second
SPOP: 141843.97 requests per second
LPUSH (needed to benchmark LRANGE): 141442.72 requests per second
LRANGE_100 (first 100 elements): 48192.77 requests per second
LRANGE_300 (first 300 elements): 14330.75 requests per second
LRANGE_500 (first 450 elements): 10405.83 requests per second
LRANGE_600 (first 600 elements): 7964.95 requests per second
MSET (10 keys): 107758.62 requests per second
CPU: Intel(R) Xeon(R) CPU E5-2683 v3 @ 2.00GHz
NIC:82599ES 10-Gigabit SFI/SFP+ Network Connection (rev 01)
ANS run on a lcore.
4 dpdk-nginx run on ANS.
# ./wrk --timeout=1 --latency -H "Connection: close" -t20 -c100 -d30s http://10.0.0.2
Running 30s test @ http://10.0.0.2
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 480.80us 73.12us 4.66ms 88.86%
Req/Sec 5.32k 125.94 6.98k 84.30%
Latency Distribution
50% 478.00us
75% 505.00us
90% 535.00us
99% 648.00us
3186335 requests in 30.10s, 2.51GB read
Requests/sec: 105860.26
Transfer/sec: 85.31MB
ANS run on two lcore.
8 dpdk-nginx run on ANS.
# ./wrk --timeout=1 --latency -H "Connection: close" -t20 -c100 -d30s http://10.0.0.2
Running 30s test @ http://10.0.0.2
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 251.70us 89.45us 4.39ms 68.59%
Req/Sec 10.28k 340.17 12.40k 67.67%
Latency Distribution
50% 246.00us
75% 310.00us
90% 363.00us
99% 480.00us
6155775 requests in 30.10s, 4.84GB read
Requests/sec: 204512.88
Transfer/sec: 164.81MB
CPU: Intel(R) Xeon(R) CPU E5-2683 v3 @ 2.00GHz
NIC:82599ES 10-Gigabit SFI/SFP+ Network Connection (rev 01)
ANS run on a lcore.
8 dpdk-nginx run on ANS.
# ./wrk --timeout=1 --latency -t20 -c100 -d30s http://10.0.0.2
Running 30s test @ http://10.0.0.2
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 257.91us 407.36us 24.37ms 98.77%
Req/Sec 20.85k 1.84k 26.88k 70.23%
Latency Distribution
50% 214.00us
75% 289.00us
90% 338.00us
99% 825.00us
12488349 requests in 30.10s, 9.89GB read
Requests/sec: 414900.42
Transfer/sec: 336.31MB
ANS run on two lcore.
10 dpdk-nginx run on ANS.
# ./wrk --timeout=1 --latency -t20 -c100 -d30s http://10.0.0.2
Running 30s test @ http://10.0.0.2
20 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 184.60us 165.39us 12.97ms 98.92%
Req/Sec 26.69k 0.93k 29.92k 79.17%
Latency Distribution
50% 186.00us
75% 200.00us
90% 216.00us
99% 370.00us
15985217 requests in 30.10s, 12.65GB read
Requests/sec: 531077.97
Transfer/sec: 430.48MB
Examples
You can get more information and instructions from wiki page.
Notes
EAL: WARNING: Address Space Layout Randomization (ASLR) is enabled in the kernel.
EAL: This may cause issues with mapping memory into secondary processes
$ sudo sysctl -w kernel.randomize_va_space=0
You shall modify the NIC configuration in ans_main.c based on your NIC type.
ANS didn't support loopback interface, so socket client and server can't be in the same ANS tcp/ip stack.
In order to improve ANS performance, you shall isolate ANS'lcore from kernel by isolcpus and isolcate interrupt from ANS's lcore by update /proc/irq/default_smp_affinity file.
ANS run as dpdk primary process, when startup ANS, shall stop other secondary processes(nginx/redis/http_server).
Don't run ANS on lcore0, it will effect ANS performance.
You shall include dpdk libs as below way because mempool lib has attribute((constructor, used)) in latest version, otherwise your application would coredump.
-$(RTE_ANS)/librte_anssock/librte_anssock.a \
-L$(RTE_SDK)/$(RTE_TARGET)/lib \
-Wl,--whole-archive -Wl,-lrte_mbuf -Wl,-lrte_mempool_ring -Wl,-lrte_mempool -Wl,-lrte_ring -Wl,-lrte_eal -Wl,--no-whole-archive -Wl,-export-dynamic -lnuma \
Support
BSD LICENSE, you may use ANS freely.
For free support, please use ANS team mail list at anssupport@163.com, or QQ Group:86883521 (FULL), 519855957(No Full), or https://dpdk-ans.slack.com, or skype(bluenet13).