memslap - Load testing and benchmarking tool for memcached
memslap is a load generation and benchmark tool for memcached(1) servers. It generates configurable workload such as threads, concurrencies, connections, run time, overwrite, miss rate, key size, value size, get/set proportion, expected throughput, and so on. Furthermore, it also supports data verification, expire-time verification, UDP, binary protocol, facebook test, replication test, multi-get and reconnection, etc.
Memslap manages network connections like memcached with libevent. Each thread of memslap is bound with a CPU core, all the threads don't communicate with each other, and there are several socket connections in each thread. Each connection keeps key size distribution, value size distribution, and command distribution by itself.
You can specify servers via the --servers option or via the environment variable
Memslap is developed to for the following purposes:
For memslap, both TCP and UDP use non-blocking network IO. All the network events are managed by libevent as memcached. The network module of memslap is similar to memcached. Libevent can ensure memslap can handle network very efficiently.
Memslap has the similar implementation of multi-threads to memcached. Memslap creates one or more self-governed threads; each thread is bound with one CPU core if the system supports setting CPU core affinity.
In addition, each thread has a libevent to manage the events of the network; each thread has one or more self-governed concurrencies; and each concurrency has one or more socket connections. All the concurrencies don’t communicate with each other even though they are in the same thread.
Memslap can create thousands of socket connections, and each concurrency has tens of socket connections. Each concurrency randomly or sequentially selects one socket connection from its socket connection pool to run, so memslap can ensure each concurrency handles one socket connection at any given time. Users can specify the number of concurrency and socket connections of each concurrency according to their expected workload.
In order to improve time efficiency and space efficiency, memslap creates a random characters table with 10M characters. All the suffixes of keys and values are generated from this random characters table.
Memslap uses the offset in the character table and the length of the string to identify a string. It can save much memory. Each key contains two parts, a prefix and a suffix. The prefix is an uint64_t, 8 bytes. In order to verify the data set before, memslap need to ensure each key is unique, so it uses the prefix to identify a key. The prefix cannot include illegal characters, such as ‘\r’, ‘\n’, ‘\0’ and ‘ ‘. And memslap has an algorithm to ensure that.
Memslap doesn’t generate all the objects (key-value pairs) at the beginning. It only generates enough objects to fill the task window (default 10K objects) of each concurrency. Each object has the following basic information, key prefix, key suffix offset in the character table, key length, value offset in the character table, and value length.
In the work process, each concurrency sequentially or randomly selects an object from the window to do set operation or get operation. At the same time, each concurrency kicks objects out of its window and adds new object into it.
Memslap uses libevent to schedule all the concurrencies of threads, and each concurrency schedules tasks based on the local task window. Memslap assumes that if each concurrency keeps the same key distribution, value distribution and commands distribution, from outside, memslap keeps all the distribution as a whole. Each task window includes a lot of objects, each object stores its basic information, such as key, value, expire time, and so on. At any time, all the objects in the window keep the same and fixed key and value distribution. If an object is overwritten, the value of the object will be updated. Memslap verifies the data or expire-time according to the object information stored in the task window.
Libevent selects which concurrency to handle based on a specific network event. Then the concurrency selects which command (get or set) to operate based on the command distribution. If it needs to kick out an old object and add a new object, in order to keep the same key and value distribution, the new object must have the same key length and value length.
If memcached server has two cache layers (memory and SSD), running memslap with different window sizes can get different cache miss rates. If memslap adds enough objects into the windows at the beginning, and the cache of memcached cannot store all the objects initialized, then memslap will get some objects from the second cache layer. It causes the first cache layer to miss. So the user can specify the window size to get the expected miss rate of the first cache layer.
Because each thread is self-governed, memslap can assign different threads to handle different memcached servers. This is just one of the ways in which memslap supports multiple servers. The only limitation is that the number of servers cannot be greater than the number of threads. The other way to support multiple servers is for replication test. Each concurrency has one socket connection to each memcached server. For the implementation, memslap can set some objects to one memcached server, and get these objects from the other servers.
By default, Memslap does single get. If the user specifies multi-get option, memslap will collect enough get commands and pack and send the commands together.
Memslap supports both the ASCII protocol and binary protocol, but it runs on the ASCII protocol by default. Memslap by default runs on the TCP protocol, but it also supports UDP. Because UDP is unreliable, dropped packages and out-of-order packages may occur. Memslap creates a memory buffer to handle these problems. Memslap tries to read all the response data of one command from the server and reorders the response data. If some packages get lost, the waiting timeout mechanism can ensure half-baked packages will be discarded and the next command will be sent.
Below are some usage samples:
The user must specify one server at least to run memslap. The rest of the parameters have default values, as shown below:
Thread number = 1 Concurrency = 16
Run time = 600 seconds Configuration file = NULL
Key size = 64 Value size = 1024
Get/set = 9:1 Window size = 10k
Execute number = 0 Single get = true
Multi-get = false Number of sockets of each concurrency = 1
Reconnect = false Data verification = false
Expire-time verification = false ASCII protocol = true
Binary protocol = false Dumping statistic information
periodically = false
Overwrite proportion = 0% UDP = false
TCP = true Limit throughput = false
Facebook test = false Replication test = false
All the distributions are read from the configuration file specified by user with “—cfg_cmd” option. If the user does not specify a configuration file, memslap will run with the default distribution (key size = 64, value size = 1024, get/set = 9:1). For information on how to edit the configuration file, refer to the “Configuration File” section.
The minimum key size is 16 bytes; the maximum key size is 250 bytes. The precision of proportion is 0.001. The proportion of distribution will be rounded to 3 decimal places.
The minimum value size is 1 bytes; the maximum value size is 1M bytes. The precision of proportion is 0.001. The proportion of distribution will be rounded to 3 decimal places. Currently, memslap only supports set and get commands. And it supports 100% set and 100% get. For 100% get, it will preset some objects to the server.
The high performance of memslap benefits from the special schedule of thread and concurrency. It’s important to specify the proper number of them. The default number of threads is 1; the default number of concurrency is 16. The user can use “—threads” and “--concurrency” to specify these variables.
If the system supports setting CPU affinity and the number of threads specified by the user is greater than 1, memslap will try to bind each thread to a different CPU core. So if you want to get the best performance memslap, it is better to specify the number of thread equal to the number of CPU cores. The number of threads specified by the user can also be less or greater than the number of CPU cores. Because of the limitation of implementation, the number of concurrencies could be the multiple of the number of threads.
1. For 8 CPU cores system
2. For 16 CPU cores system
The memslap performs very well, when used to test the performance of memcached servers. Most of the time, the bottleneck is the network or the server. If for some reason the user wants to limit the performance of memslap, there are two ways to do this:
Decrease the number of threads and concurrencies. Use the option “--tps” that memslap provides to limit the throughput. This option allows the user to get the expected throughput. For example, assume that the maximum throughput is 50 kops/s for a specific configuration, you can specify the throughput equal to or less than the maximum throughput using “--tps” option.
Most of the time, the user does not need to specify the window size. The default window size is 10k. For Schooner Memcached, the user can specify different window sizes to get different cache miss rates based on the test case. Memslap supports cache miss rate between 0% and 100%. If you use this utility to test the performance of Schooner Memcached, you can specify a proper window size to get the expected cache miss rate. The formula for calculating window size is as follows:
Assume that the key size is 128 bytes, and the value size is 2048 bytes, and concurrency=128.
1. Small cache cache_size=1M, 100% cache miss (all data get from SSD). win_size=10k
(1). cache miss rate 0%
(2). cache miss rate 5%
(1). cache miss rate 0%
(2). cache miss
The formula for calculating window size for cache miss rate 0%:
cache_size / concurrency / (key_size + value_size) * 0.5
The formula for calculating window size for cache miss rate 5%:
cache_size / concurrency / (key_size + value_size) * 0.7
Memslap supports both data verification and expire-time verification. The user can use "--verify=" or "-v" to specify the proportion of data verification. In theory, it supports 100% data verification. The user can use "--exp_verify=" or "-e" to specify the proportion of expire-time verification. In theory, it supports 100% expire-time verification. Specify the "--verbose" options to get more detailed error information.
For example: --exp_verify=0.01 –verify=0.1 , it means that 1% of the objects set with expire-time, 10% of the objects gotten will be verified. If the objects are gotten, memslap will verify the expire-time and value.
Memslap supports multi-servers based on self-governed thread. There is a limitation that the number of servers cannot be greater than the number of threads. Memslap assigns one thread to handle one server at least. The user can use the "--servers=" or "-s" option to specify multi-servers.
--servers=10.1.1.1:11211,10.1.1.2:11212,10.1.1.3:11213 --threads=6 --concurrency=36
The above command means that there are 6 threads, with each thread having 6 concurrencies and that threads 0 and 3 handle server 0 (10.1.1.1); threads 1 and 4 handle server 1 (10.1.1.2); and thread 2 and 5 handle server 2 (10.1.1.3).
All the threads and concurrencies in memslap are self-governed.
So is memslap. The user can start up several memslap instances. The user can run memslap on different client machines to communicate with the same memcached server at the same. It is recommended that the user start different memslap on different machines using the same configuration.
The default memslap runs with time mode. The default run time is 10 minutes. If it times out, memslap will exit. Do not specify both execute number mode and time mode at the same time; just specify one instead.
--time=30s (It means the test will run 30 seconds.)
--execute_number=100000 (It means that after running 100000 commands, the test will exit.)
The user can use "--stat_freq=" or "-S" to specify the frequency.
Memslap will dump the statistics of the commands (get and set) at the frequency of every 20 seconds.
For more information on the format of dumping statistic information, refer to “Format of Output” section.
The user can use "--division=" or "-d" to specify multi-get keys count. Memslap by default does single get with TCP. Memslap also supports data verification and expire-time verification for multi-get.
Memslap supports multi-get with both TCP and UDP. Because of the different implementation of the ASCII protocol and binary protocol, there are some differences between the two. For the ASCII protocol, memslap sends one “multi-get” to the server once. For the binary protocol, memslap sends several single get commands together as “multi-get” to the server.
Memslap supports both UDP and TCP. For TCP, memslap does not reconnect the memcached server if socket connections are lost. If all the socket connections are lost or memcached server crashes, memslap will exit. If the user specifies the “--reconnect” option when socket connections are lost, it will reconnect them.
User can use “--udp” to enable the UDP feature, but UDP comes with some limitations:
UDP cannot set data more than 1400 bytes.
UDP is not supported by the binary protocol because the binary protocol of memcached does not support that.
UDP doesn’t support reconnection.
Set data with TCP and multi-get with UDP. Specify the following options:
If you want to create thousands of TCP connections, specify the
For example: --facebook --division=50 --conn_sock=200
The above command means that memslap will do facebook test, each concurrency has 200 socket TCP connections and one UDP socket.
Memslap sets objects with the TCP socket, and multi-gets 50 objects once with the UDP socket.
If you specify "--division=50", the key size must be less that 25 bytes because the UDP packet size is 1400 bytes.
For replication test, the user must specify at least two memcached servers. The user can use “—rep_write=” option to enable feature.
The above command means that there are 2 replication memcached servers, memslap will set objects to both server 0 and server 1, get objects which are set to server 0 before from server 1, and also get objects which are set to server 1 before from server 0. If server 0 crashes, memslap will only get objects from server 1. If server 0 comes back to life again, memslap will reconnect server 0. If both server 0 and server 1 crash, memslap will exit.
Start memslap with "--conn_sock=" or "-n" to enable this feature. Make sure that your system can support opening thousands of files and creating thousands of sockets. However, this feature does not support reconnection if sockets disconnect.
--threads=8 --concurrency=128 --conn_sock=128
The above command means that memslap starts up 8 threads, each thread has 16 concurrencies, each concurrency has 128 TCP socket connections, and the total number of TCP socket connections is 128 * 128 = 16384.
Start memslap with "--binary" or "-B" options to enable this feature. It supports all the above features except UDP, because the latest memcached 1.3.3 does not implement binary UDP protocol.
Since memcached 1.3.3 doesn't implement binary UDP protocol, memslap does not support UDP. In addition, memcached 1.3.3 does not support multi-get. If you specify "--division=50" option, it just sends 50 get commands together as “mulit-get” to the server.
This section describes the format of the configuration file. By default when no configuration file is specified memslap reads the default one located at ~/.memslap.cnf.
Below is a sample configuration file:
*************************************************************************** #comments should start with '#' #key #start_len end_len proportion # #key length range from start_len to end_len #start_len must be equal to or greater than 16 #end_len must be equal to or less than 250 #start_len must be equal to or greater than end_len #memslap will generate keys according to the key range #proportion: indicates keys generated from one range accounts for the total generated keys # #example1: key range 16~100 accounts for 80% # key range 101~200 accounts for 10% # key range 201~250 accounts for 10% # total should be 1 (0.8+0.1+0.1 = 1) # # 16 100 0.8 # 101 200 0.1 # 201 249 0.1 # #example2: all keys length are 128 bytes # # 128 128 1 key 128 128 1 #value #start_len end_len proportion # #value length range from start_len to end_len #start_len must be equal to or greater than 1 #end_len must be equal to or less than 1M #start_len must be equal to or greater than end_len #memslap will generate values according to the value range #proportion: indicates values generated from one range accounts for the total generated values # #example1: value range 1~1000 accounts for 80% # value range 1001~10000 accounts for 10% # value range 10001~100000 accounts for 10% # total should be 1 (0.8+0.1+0.1 = 1) # # 1 1000 0.8 # 1001 10000 0.1 # 10001 100000 0.1 # #example2: all value length are 128 bytes # # 128 128 1 value 2048 2048 1 #cmd #cmd_type cmd_proportion # #currently memslap only supports get and set command. # #cmd_type #set 0 #get 1 # #example: set command accounts for 50% # get command accounts for 50% # total should be 1 (0.5+0.5 = 1) # # cmd # 0 0.5 # 1 0.5 cmd 0 0.1 1.0 0.9
At the beginning, memslap displays some configuration information as follows:
The servers used by memslap.
The number of threads memslap runs with.
The number of concurrencies memslap runs with.
How long to run memslap.
The task window size of each concurrency.
The proportion of set command.
The proportion of get command.
The output of dynamic statistics is something like this:
--------------------------------------------------------------------------------------------------------------------------------- Get Statistics Type Time(s) Ops TPS(ops/s) Net(M/s) Get_miss Min(us) Max(us) Avg(us) Std_dev Geo_dist Period 5 345826 69165 65.3 0 27 2198 203 95.43 177.29 Global 20 1257935 62896 71.8 0 26 3791 224 117.79 192.60 Set Statistics Type Time(s) Ops TPS(ops/s) Net(M/s) Get_miss Min(us) Max(us) Avg(us) Std_dev Geo_dist Period 5 38425 7685 7.3 0 42 628 240 88.05 220.21 Global 20 139780 6989 8.0 0 37 3790 253 117.93 224.83 Total Statistics Type Time(s) Ops TPS(ops/s) Net(M/s) Get_miss Min(us) Max(us) Avg(us) Std_dev Geo_dist Period 5 384252 76850 72.5 0 27 2198 207 94.72 181.18 Global 20 1397720 69886 79.7 0 26 3791 227 117.93 195.60 ---------------------------------------------------------------------------------------------------------------------------------
Statistics information of get command
Statistics information of set command
Statistics information of both get and set command
Result within a period
The rate of network
How many objects can’t be gotten
The minimum response time
The maximum response time
The average response time
Standard deviation of response time
Geometric distribution based on natural exponential function
At the end, memslap will output something like this:
--------------------------------------------------------------------------------------------------------------------------------- Get Statistics (1257956 events) Min: 26 Max: 3791 Avg: 224 Geo: 192.60 Std: 116.23 Log2 Dist: 4: 0 10 84490 215345 8: 484890 459823 12543 824 12: 31 Set Statistics (139782 events) Min: 37 Max: 3790 Avg: 253 Geo: 224.84 Std: 116.83 Log2 Dist: 4: 0 0 4200 16988 8: 50784 65574 2064 167 12: 5 Total Statistics (1397738 events) Min: 26 Max: 3791 Avg: 227 Geo: 195.60 Std: 116.60 Log2 Dist: 4: 0 10 88690 232333 8: 535674 525397 14607 991 12: 36 cmd_get: 1257969 cmd_set: 139785 get_misses: 0 verify_misses: 0 verify_failed: 0 expired_get: 0 unexpired_unget: 0 written_bytes: 242516030 read_bytes: 1003702556 object_bytes: 152086080 packet_disorder: 0 packet_drop: 0 udp_timeout: 0 Run time: 20.0s Ops: 1397754 TPS: 69817 Net_rate: 59.4M/s ---------------------------------------------------------------------------------------------------------------------------------
Get statistics of response time
Set statistics of response time
Both get and set statistics of response time
The accumulated and minimum response time
The accumulated and maximum response time
The accumulated and average response time
Standard deviation of response time
Geometric distribution based on logarithm 2
Total get commands done
Total set commands done
How many objects can’t be gotten from server
How many objects need to verify but can’t get them
How many objects with insistent value
How many objects are expired but we get them
How many objects are unexpired but we can’t get them
Total written bytes
Total read bytes
Total object bytes
How many UDP packages are disorder
How many UDP packages are lost
How many times UDP time out happen
Total run time
The average rate of network
-s, --servers= List one or more servers to connect. Servers count must be less than threads count. e.g.: --servers=localhost:1234,localhost:11211
-T, --threads= Number of threads to startup, better equal to CPU numbers. Default 8.
-c, --concurrency= Number of concurrency to simulate with load. Default 128.
-n, --conn_sock= Number of TCP socks per concurrency. Default 1.
-x, --execute_number= Number of operations(get and set) to execute for the given test. Default 1000000.
-t, --time= How long the test to run, suffix: s-seconds, m-minutes, h-hours, d-days e.g.: --time=2h.
-F, --cfg_cmd= Load the configure file to get command,key and value distribution list.
-w, --win_size= Task window size of each concurrency, suffix: K, M e.g.: --win_size=10k. Default 10k.
-X, --fixed_size= Fixed length of value.
-v, --verify= The proportion of date verification, e.g.: --verify=0.01
-d, --division= Number of keys to multi-get once. Default 1, means single get.
-S, --stat_freq= Frequency of dumping statistic information. suffix: s-seconds, m-minutes, e.g.: --resp_freq=10s.
-e, --exp_verify= The proportion of objects with expire time, e.g.: --exp_verify=0.01. Default no object with expire time
-o, --overwrite= The proportion of objects need overwrite, e.g.: --overwrite=0.01. Default never overwrite object.
-R, --reconnect Reconnect support, when connection is closed it will be reconnected.
-U, --udp UDP support, default memslap uses TCP, TCP port and UDP port of server must be same.
-a, --facebook Whether it enables facebook test feature, set with TCP and multi-get with UDP.
-B, --binary Whether it enables binary protocol. Default with ASCII protocol.
-P, --tps= Expected throughput, suffix: K, e.g.: --tps=10k.
-p, --rep_write= The first nth servers can write data, e.g.: --rep_write=2.
-b, --verbose Whether it outputs detailed information when verification fails.
-h, --help Display this message and then exit.
-V, --version Display the version of the application and then exit.
memslap -s 127.0.0.1:11211 -S 5s
memslap -s 127.0.0.1:11211 -t 2m -v 0.2 -e 0.05 -b
memslap -s 127.0.0.1:11211 -F config -t 2m -w 40k -S 20s -o 0.2
memslap -s 127.0.0.1:11211 -F config -t 2m -T 4 -c 128 -d 20 -P 40k
memslap -s 127.0.0.1:11211 -F config -t 2m -d 50 -a -n 40
memslap -s 127.0.0.1:11211,127.0.0.1:11212 -F config -t 2m
memslap -s 127.0.0.1:11211,127.0.0.1:11212 -F config -t 2m -p 2
To find out more information please check: http://launchpad.org/libmemcached
Mingqiang Zhuang <firstname.lastname@example.org> (Schooner Technolgy) Brian Aker, <email@example.com>