dperf

English | 中文

dperf is a DPDK based 100Gbps network performance and load testing software.

Advantage

High performance：
- Based on DPDK, dperf can generate huge traffic with a single x86 server: tens of millions of HTTP CPS，hundreds of Gbps throughput and billions of concurrent connections.
Detailed statistics：
- Provides detailed statistics and identify every packet loss.
Support multiple scenarios：
- Load testing and stability testing for Layer 4 Load Balancer and other Layer 4 gateways.
- Network performance testing for servers on cloud.
- Performance test of network package processing capability for NIC and CPU.
- Can be used as a high performance HTTP server or client for load testing.

Performance

HTTP Connections per Second

Client Cores	Server Cores	HTTP CPS
1	1	2,101,044
2	2	4,000,423
4	4	7,010,743
6	6	10,027,172

HTTP Throughput per Second

Client Cores	Server Cores	RX(Gbps)	TX(Gbps)	Client CPU Usage(%)	Server CPU Usage(%)
1	1	18	18	60	59
2	2	35	35	60	59
4	4	46	46	43	43

HTTP Current Connections

Client Cores	Server Cores	Current Connections	Client CPU Usage(%)	Server CPU Usage(%)
1	1	100,000,000	34	39
2	2	200,000,000	36	39
4	4	400,000,000	40	41

UDP TX PPS

Client Cores	TX MPPS	Client CPU Usage(%)
1	15.96	95
2	29.95	95
4	34.92	67
6	35.92	54
8	37.12	22

Note: this test is under one 25Gbps Mellanox CX4

Client & Server Configuration

Above performance is obtained with below configurations:

MEM: 512GB(hugepage 100GB)
NIC: Mellanox MT27710 25Gbps * 2
Kernel: 4.19.90

Statistics

dperf outputs various statistics every second：

TPS, CPS, various PPS
Errors of TCP/Socket/HTTP
Packets loss/drop
Retransmissions of TCP Flags

seconds 22                 cpuUsage 52
pktRx   3,001,058          pktTx    3,001,025          bitsRx   2,272,799,040      bitsTx  1,920,657,600      dropTx  0
arpRx   0                  arpTx    0                  icmpRx   0                  icmpTx  0                  otherRx 0          badRx 0
synRx   1,000,345          synTx    1,000,330          finRx    1,000,350          finTx   1,000,350          rstRx   0          rstTx 0
synRt   0                  finRt    0                  ackRt    0                  pushRt  0                  tcpDrop 0
skOpen  1,000,330          skClose  1,000,363          skCon    230                skErr   0
httpGet 1,000,345          http2XX  1,000,350          httpErr  0
ierrors 0                  oerrors  0                  imissed  0

Getting Started

Set hugepages

#edit '/boot/grub2/grub.cfg' like this, and reboot the OS
linux16 /vmlinuz-... nopku transparent_hugepage=never default_hugepagesz=1G hugepagesz=1G hugepages=8

Build DPDK

#edit 'config/common_base' to enable PMDs
#Mellanox CX4/CX5 requires 'CONFIG_RTE_LIBRTE_MLX5_PMD=y'
#HNS3 requires 'CONFIG_RTE_LIBRTE_HNS3_PMD=y'
#VMXNET3 requires 'CONFIG_RTE_LIBRTE_VMXNET3_PMD=y'

TARGET=x86_64-native-linuxapp-gcc #or arm64-armv8a-linuxapp-gcc

cd /root/dpdk/dpdk-stable-19.11.10
make install T=$TARGET -j16

Build dperf

cd dperf
make -j8 RTE_SDK=/root/dpdk/dpdk-stable-19.11.10 RTE_TARGET=$TARGET

Bind interface

#Mellanox NIC skip this step!
#Suppose your PCI number is 0000:1b:00.0

modprobe uio
modprobe uio_pci_generic
/root/dpdk/dpdk-stable-19.11.10/usertools/dpdk-devbind.py -b uio_pci_generic 0000:1b:00.0

Start dperf server

#dperf server bind at 6.6.241.27:80,  gateway is 6.6.241.1
./build/dperf -c test/http/server-cps.conf

Send request from a client

#the client IP must be in the range of 'client' in the configuration file
ping 6.6.241.27
curl http://6.6.241.27/

Running the tests

Below example will start a HTTP CPS stress test.

#run dperf server
./build/dperf -c test/http/server-cps.conf

#from another host, run dperf client
./build/dperf -c test/http/client-cps.conf

Documentation

See the website at https://dperf.org/.

Limitation

dperf requires that the HTTP message is in one packet, which is not suitable for the test of layer 7 load balancer.
dperf requires exclusive use of the network interfaces.
dperf does not have routing capability. It is recommended to build a test environment with a switch.

Contributing

dperf welcomes your contribution. See the CONTRIBUTING file for details.

Author

Jianzhang Peng, a Ph.D. graduate in Computer Science from the University of Science and Technology of China, served as a Principal Engineer at both Huawei and Baidu. He participated in the development of 7-layer and 4-layer load balancing systems. During his tenure at Baidu, he developed the dperf project. Even after leaving Baidu, he remains dedicated to the maintenance of the dperf project. Currently, he works in Singapore, focusing on researching and implementing low-latency network systems for high-frequency crypto quantitative trading.

License

dperf is distributed under the Apache License, Version 2.0.

yuexiaocai/dperf