Apache bench (AB) stress test Overview - from 0 to 1, covering all major use scenarios

Preface

Apache Bench(ab) Mainly for HTTP Service provides performance stress testing , Here is a summary of ab Pressure measurement method used in daily use .

ab Closer to pragmatism , Like other tools, it does not include a large number of parameters with few usage scenarios , When you see man perhaps --help When helping the dense parameters displayed in the document , It is inevitable to persuade some people to retreat .

This article prefers production 、 Daily use , You won't drill through every parameter , When you read this article , It must have come with the need to solve the problem , Try to let you use the least time cost , For business performance 、 There is a roughly complete assessment of load capacity .

One 、 How to install ab

1.Linux

It is recommended to install directly from the software source , The dependencies required by each component will be installed by default , The compilation and installation requires manual installation of dependencies , The naming of the dependent components of each distribution may also be different , You may also encounter various errors on the way , It's like a recursion , You have to sink deep , So I prefer practicality and minimalism , Unless the version of the software source is too old to meet the requirements , Or look at each version Changelog Then confirm to specify an installation version , Then you can choose to compile and install . Installation is not the focus of this article , Try to pass .

2.Windows

Windows client See this link , Download the compressed file and unzip it , take bin The directory is set to environment variable or cmd Enter into bin Under the table of contents , Then use ab Command is enough .

Two 、 Parameter optimization of pressure measuring machine

When using pressure measuring tools for pressure testing , Will be limited by tool performance or machine performance , If it is limited by the hardware performance of the machine , It is recommended to use multiple clients for pressure testing , At the same time, the single machine itself also has some network related kernel parameters that can be optimized .

1.tcp_fin_timeout

MSL namely ：Maximum Segment Lifetime, It defines the maximum time that any message is allowed to survive , If it exceeds, discard .

that tcp_fin_timeout It's used to set up MSL Of , The default is 60s.

cat /proc/sys/net/ipv4/tcp_fin_timeout

2.tcp_tw_reuse

linux The number of file handles allocated to each user is limited , After disconnection , Ports and related resources will not be released immediately , But into time_wait state , wait for 2MSL Enter after CLOSE state , At this time, the connection is released , You can refer to RFC793 The definition of ：

obviously , In the pressure measurement scenario , We need to build as many pressure measuring machines as possible in a short time TCP Connect , and time_wait If the state remains 2MSL( namely 2 minute ), So this 2MSL Corresponding to time_wait Once the occupied handle reaches the upper limit , Cannot add new TCP Connect .

We need to take tcp_tw_reuse Parameters on , Let the system turn on time_wait State reuse , Allow resources in this state to be reused for new TCP Connect .

Just open it temporarily ：

sysctl -w sysctl net.ipv4.tcp_tw_reuse = 1

perhaps ：

echo 1 > /proc/sys/net/ipv4/tcp_tw_reuse

If you need to permanently open , You need to write to /etc/sysctl.conf, And perform sysctl -p take effect .

3.tcp_tw_recycle

seeing the name of a thing one thinks of its function , This parameter is used to recycle time_wait State resources , It's usually with tcp_tw_reuse In combination with .

Temporarily open ：

sysctl -w sysctl net.ipv4.tcp_tw_recycle = 1

perhaps ：

echo 1 > /proc/sys/net/ipv4/tcp_tw_recycle

If you need to permanently open , You need to write to /etc/sysctl.conf, And perform sysctl -p take effect .

4.ip_local_port_range

This parameter indicates the range of ports that can be used for local to external connections , The default value of the parameter is 32768 60999, That is, a single machine can create a total of 28231 A connection ：

According to the pressure test scenario , Adjust this parameter appropriately .

For example, increase more 1000 Ports , It can be ：

echo '32768 61999' > /proc/sys/net/ipv4/ip_local_port_range

5. Count the number of connections in each state

Last , Count the number of connections in each state , It can be ：

netstat -an|awk '/tcp/{print $6}'|sort |uniq -c
netstat -an|awk '/tcp/{s[$6]++}END{for(i in s)print i,s[i]}'

3、 ... and 、 Start the stress test journey

Use --help You can see the parameter support list ：

The pressure measurement format is ：ab [options] [http[s]]://]hostname[:port]/path/

Be careful ：options Don't put it on hostname Back , This is not supported .

1. Pressure measurement short connection (-m/-c/-n)

Use HTTP GET request , Maximum concurrency per time 1000, Pressure measurement quantity 10000 Time ,：

ab -m GET -c 1000 -n 10000  http://hostname/

Be careful ： The target must be followed by a path , If it is the root path, bring "/", Like the one above hostname/, Otherwise, an error will be reported .

paraphrase ：

Output information ：

Server Software:        nginx
Server Hostname:        xxx
Server Port:            80

Document Path:          /
Document Length:        231 bytes  # Returned page size 

Concurrency Level:      1000       # Number of concurrent connections 1000
Time taken for tests:   0.457 seconds   # Total time 
Complete requests:      10000          # Always complete the request 
Failed requests:        0              # Failed request 
Total transferred:      4560000 bytes  # Total bytes transferred 
HTML transferred:       2310000 bytes   #html Number of bytes transferred 
Requests per second:    21896.07 [#/sec] (mean)  # Average requests per second 
Time per request:       45.670 [ms] (mean        # Time per concurrent request 
Time per request:       0.046 [ms] (mean, across all concurrent requests)  # The average time each request actually runs .
Transfer rate:          9750.59 [Kbytes/sec] received  # Transmission rate 

Connection Times (ms)   # Connection processing time during stress testing 
              min  mean[+/-sd] median   max
Connect:        1   20   2.7     20      28 
Processing:     7   24   5.0     23      43
Waiting:        1   17   4.8     16      33
Total:         22   43   4.1     43      57

Percentage of the requests served within a certain time (ms) # The proportion of request response time in a certain time 
  50%     43                      # Half of the requests are completed , Statistical average response time , And so on 
  66%     44
  75%     46
  80%     47
  90%     48
  95%     50
  98%     52
  99%     55
 100%     57 (longest request)

Corresponding packet capture ：

It's not hard to see. ,ab every time GET Will shake hands with the destination , Will not reuse the same TCP Connect , Sent in a very small time 10000 Time GET request , And establish 10000 Time , produce 10000 One session .

If a larger dose is needed , Such as ：

ab -m GET -c 10000 -n 1000000  http://hostname/

-c 10000, At most 10000 A request , Because the concurrency is large , The maximum number of files opened by a single user may be exceeded , At this time, you need to set ulimit：

ulimit -n 65535

ulimit It's a kind of Linux The internal key function of the system , It has a set of parameters , Used to generate for it shell Set resource usage restrictions for processes and their children . Set user open files（ The maximum number of files that a user can open ）.

View the number of files opened by the current user , You can use the following command ：

lsof -u `whoami` |wc -l

2. Pressure gauge connection (-k)

ab -m GET -k -c 1000 -n 100000 http://hostname/

It is obvious that the long connection mode pressure measurement ,QPS There is a significant improvement , Because every time GET Resources try to reuse a connection , Unlike short connections, each time GET Shake hands before GET.

Simultaneously from HTTP Header field Connetction:keep-alive It can be seen that -k The parameter pressure measurement mode is in HTTP Add this field to the header ：

Why is there no timeout or significantly longer response time at the opposite end during the pressure test ？

To say the least , At this time, the client request level <= Server business layer processing level , So only one client is used to test the server , If the opposite end does not reach the performance bottleneck , Then you can consider increasing the pressure measurement concurrency or multiple clients' pressure measurement at the same time .

Below is an Internet web The tester is for display ：

The above figure shows that the server is already in high load mode , Unable to respond to the request normally and return 502 Status code .

Use htop Observe the load and memory occupation of the server ,CPU、 Memory is full ：

Count the connections established by the server ：

lsof -i :80|awk '/ESTABLISHED/{S+=1}END{print S}'

this 1000 Connections are ab -c 1000 The specified quantity .

If you want to specify statistics IP Connections that have been established , It can be ：

lsof -i :80|awk '/ESTABLISHED/&&/ipaddress/{S+=1}END{print S}'

3. Specify the timeout period (-s)

-s s Used to specify the timeout , Unit second . from man The help document shows that this parameter is in 2.4.4 Later versions are available , And the default is 30s Overtime .

It's easy to use , It can be used together with other pressure measurement parameters .

In long connection mode , Every time GET The most requested 5s Treat as a timeout request ：

ab -m GET -s 3 -k -c 1000 -n 10000 http://hostname/

In short connection mode , Every time POST The most requested 5s Treat as a timeout request ：

ab -m POST -s 5 -c 1000 -n 10000 'http://hostname/?username=xxx&passwd=xxx'

4. Appoint TLS Version pressure test (-f)

-f Appoint TLS Protocol version ：

How to know what protocol versions are supported on the opposite side ？

openssl s_client -connect  hostname:443 -tls1/-tls1_2/-tls1_3  # Notice the "/" You can't write it directly , Only one at a time tls edition , More than one can be traversed by loop 

Or use Online website Query the supported TLS Protocol version .

Example ：

Appoint TLS1.1 edition ：

ab -f TLS1.1 -m GET -k -c 1000 -n 10000 https://hostname/ 

You can clearly see , The client is in TLS The handshake package declares to the server that it uses TLS1.1 edition ：

HTTPS scenario , Different versions TLS、cipher Kit , The encryption and decryption strength is inconsistent , The performance of the loss will also be inconsistent , Therefore, it will also be intuitively reflected in QPS On .

5. Insert cookies(-C)

-C Is used to specify the cookies Information , The format is similar to key value pairs ：

When it is necessary to perform pressure test in some scenarios that require login , By inserting cookies To maintain login status ：

ab -m GET -k -c 500 -n 2000 -C 'cookie-name=value'  http://hostname/

By inserting cookies, You can test any resources that can only be accessed after login .

6. Specify proxy services (-X)

Pressure test through agents , Format : -X proxy[:port]

ab -m GET -X proxyip:port -k -c 500 -n 2000 http://hostname/

Be similar to nginx Seven layer reverse proxy , The client request is sent to the proxy server , The proxy server requests the real server and returns it to the client .

7. Set up HTTP Head (-H)

-H Customize HTTP Head content

Appoint HOST：

ab -m GET -H 'Host:test.com' -c 1000 -n 10000000 -k http://hostname/

At the same time specified HOST、UA Information, etc ：

ab -m GET -H 'Host:domain' -H 'User-Agent: RokasYang/1.0 -c 1000 -n 10000000 -k http://hostname/

whatever HTTP The head can pass -H Parameter assignment .

8. Specify the total pressure measurement time (-t)

-t Used to limit the total duration of pressure measurement , Be similar to timeout command , The unit is in seconds , If it is not specified by default, there is no time limit .

Limit the maximum pressure measurement 5s end ：

ab -m GET -t 5 -c 1000 -n 10000000 http://hostname/

You can also use capinfos Command to view the time of the first and last packets in the message , interval 5s：

9. use HAED Ask for a replacement GET(-i)

This parameter can also be used -m HEAD To replace , The effect is the same , All hair HEAD request .

ab -i -k -c 1000 -n 10000000 http://hostname/

10. output to a file (-g/-e)

-g Output the measured value of each result to a file .

-e Record all requests in the percentage progress bar (1%-100%) Time consumed in , It's usually followed by csv Format file , namely Excel Format file .

Record the time of each request ：

ab -k -c 1000 -n 10000 -g output http://hostname/

therefore 1 Ten thousand requests will record ten thousand requests ( Remove the first line banner Information ).

Record the elapsed time of all requests under different progress ：

ab -k -c 1000 -n 10000 -e output.csv http://hostname/

Four 、 summary

If it is necessary to reflect that the business side tends to the most real or the highest bottleneck QPS performance , Long connection scenario is more applicable , Multiple requests reuse one TCP Connect , It saves a lot of short connections and new connections , Because of this , A long connection QPS It will always be higher than the short connection .

If the number of clients is small , The end-to-end business can carry the response normally without any 5XX HTTP State code words , You can increase the number of clients for pressure measurement , Want to grasp the most authentic service QPS performance , We need to find out whether the business is normal or abnormal / Unresponsive QPS critical point .

If you use different clients to test the same server ,QPS uneven , It's normal , When the hardware performance of different clients is different , The performance of the contract is limited , And QPS It is also closely related to network delay , Even a few ms The gap between , It may also bring multiple level QPS disparity , Especially in the small bag scenario , Each connection requires a lot of packets to interact , In this scenario, the delay gap brings QPS The gap should not be small .

Incidental PDF edition ：

Apache Bench Overview of pressure measurement scheme

Apache Bench Overview of pressure measurement scheme ( Bright color plate )