Libmemcached documentation
memaslap [options]
--help
MEMCACHED_SERVERS
memaslap is a load generation and benchmark tool for memcached 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 testss data verification, expire-time verification, UDP, binary protocol, facebook test, replication test, multi-get and reconnection, etc.
Memaslap manages network connections like memcached with libevent. Each thread of memaslap 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 memslap --servers option or via the environment variable MEMCACHED_SERVERS.
Memslap is developed to for the following purposes:
Manages network connections with libevent asynchronously.
Set both TCP and UDP up to use non-blocking IO.
Improves parallelism: higher performance in multi-threads environments.
Improves time efficiency: faster processing speed.
Generates key and value more efficiently; key size distribution and value size distribution are configurable.
Supports get, multi-get, and set commands; command distribution is configurable.
Supports controllable miss rate and overwrite rate.
Supports data and expire-time verification.
Supports dumping statistic information periodically.
Supports thousands of TCP connections.
Supports binary protocol.
Supports facebook test (set with TCP and multi-get with UDP) and replication test.
For memaslap, both TCP and UDP use non-blocking network IO. All the network events are managed by libevent as memcached. The network module of memaslap is similar to memcached. Libevent can ensure memaslap 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 testss 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 memaslap 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, memaslap 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, memaslap 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 memaslap 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, memaslap 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 memaslap with different window sizes can get different cache miss rates. If memaslap adds enough objects into the windows at the beginning, and the cache of memcached cannot store all the objects initialized, then memaslap 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, memaslap can assign different threads to handle different memcached servers. This is just one of the ways in which memaslap tests multiple servers. The only limitation is that the number of servers cannot be greater than the number of threads. The other way to test multiple servers is for replication test. Each concurrency has one socket connection to each memcached server. For the implementation, memaslap 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, memaslap will collect enough get commands and pack and send the commands together.
Memslap testss 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 tests 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:
memaslap -s 127.0.0.1:11211 -S 5s
memaslap -s 127.0.0.1:11211 -t 2m -v 0.2 -e 0.05 -b
memaslap -s 127.0.0.1:11211 -F config -t 2m -w 40k -S 20s -o 0.2
memaslap -s 127.0.0.1:11211 -F config -t 2m -T 4 -c 128 -d 20 -P 40k
memaslap -s 127.0.0.1:11211 -F config -t 2m -d 50 -a -n 40
memaslap -s 127.0.0.1:11211,127.0.0.1:11212 -F config -t 2m
memaslap -s 127.0.0.1:11211,127.0.0.1:11212 -F config -t 2m -p 2
The user must specify one server at least to run memaslap. 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, memaslap 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, memaslap only testss set and get commands. And it testss 100% set and 100% get. For 100% get, it will preset some objects to the server.
The high performance of memaslap 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 tests setting CPU affinity and the number of threads specified by the user is greater than 1, memaslap will try to bind each thread to a different CPU core. So if you want to get the best performance memaslap, 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.
For 8 CPU cores system
For example:
--threads=2 --concurrency=128
--threads=8 --concurrency=128
--threads=8 --concurrency=256
--threads=12 --concurrency=144
For 16 CPU cores system
For example:
--threads=8 --concurrency=128
--threads=16 --concurrency=256
--threads=16 --concurrency=512
--threads=24 --concurrency=288
The memaslap 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 memaslap, there are two ways to do this:
Decrease the number of threads and concurrencies. Use the option “--tps” that memaslap 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 testss 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
cache_size=4G
(1). cache miss rate 0%
win_size=8k
(2). cache miss rate 5%
win_size=11k
cache_size=16G
(1). cache miss rate 0%
win_size=32k
(2). cache miss
rate 5%
win_size=46k
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 testss both data verification and expire-time verification. The user can use "--verify=" or "-v" to specify the proportion of data verification. In theory, it testss 100% data verification. The user can use "--exp_verify=" or "-e" to specify the proportion of expire-time verification. In theory, it testss 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, memaslap will verify the expire-time and value.
Memslap testss 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.
For example:
--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 memaslap are self-governed.
So is memaslap. The user can start up several memaslap instances. The user can run memaslap on different client machines to communicate with the same memcached server at the same. It is recommended that the user start different memaslap on different machines using the same configuration.
The default memaslap runs with time mode. The default run time is 10 minutes. If it times out, memaslap will exit. Do not specify both execute number mode and time mode at the same time; just specify one instead.
For example:
--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.
For example:
--stat_freq=20s
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 testss data verification and expire-time verification for multi-get.
Memslap testss 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, memaslap sends one “multi-get” to the server once. For the binary protocol, memaslap sends several single get commands together as “multi-get” to the server.
Memslap testss both UDP and TCP. For TCP, memaslap does not reconnect the memcached server if socket connections are lost. If all the socket connections are lost or memcached server crashes, memaslap 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 testsed by the binary protocol because the binary protocol of memcached does not tests that.
UDP doesn’t tests reconnection.
Set data with TCP and multi-get with UDP. Specify the following options:
"--facebook --division=50"
If you want to create thousands of TCP connections, specify the
"--conn_sock=" option.
For example: --facebook --division=50 --conn_sock=200
The above command means that memaslap 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.
For example:
--servers=10.1.1.1:11211,10.1.1.2:11212 –rep_write=2
The above command means that there are 2 replication memcached servers, memaslap 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, memaslap will only get objects from server 1. If server 0 comes back to life again, memaslap will reconnect server 0. If both server 0 and server 1 crash, memaslap will exit.
Start memaslap with "--conn_sock=" or "-n" to enable this feature. Make sure that your system can tests opening thousands of files and creating thousands of sockets. However, this feature does not tests reconnection if sockets disconnect.
For example:
--threads=8 --concurrency=128 --conn_sock=128
The above command means that memaslap 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 memaslap with "--binary" or "-B" options to enable this feature. It testss all the above features except UDP, because the latest memcached 1.3.3 does not implement binary UDP protocol.
For example:
--binary
Since memcached 1.3.3 doesn't implement binary UDP protocol, memaslap does not tests UDP. In addition, memcached 1.3.3 does not tests 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 memaslap reads the default one located at ~/.memaslap.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 #memaslap 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 #memaslap 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 memaslap only testss 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, memaslap displays some configuration information as follows:
servers : 127.0.0.1:11211
threads count: 1
concurrency: 16
run time: 20s
windows size: 10k
set proportion: set_prop=0.10
get proportion: get_prop=0.90
servers : "servers" The servers used by memaslap.
threads count The number of threads memaslap runs with.
concurrency The number of concurrencies memaslap runs with.
run time How long to run memaslap.
windows size The task window size of each concurrency.
set proportion The proportion of set command.
get proportion 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 ---------------------------------------------------------------------------------------------------------------------------------
Get Statistics Statistics information of get command
Set Statistics Statistics information of set command
Total Statistics Statistics information of both get and set command
Period Result within a period
Global Accumulated results
Ops Total operations
TPS Throughput, operations/second
Net The rate of network
Get_miss How many objects can’t be gotten
Min The minimum response time
Max The maximum response time
Avg: The average response time
Std_dev Standard deviation of response time
Geo_dist Geometric distribution based on natural exponential function
At the end, memaslap 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 Get statistics of response time
Set Statistics Set statistics of response time
Total Statistics Both get and set statistics of response time
Min The accumulated and minimum response time
Max The accumulated and maximum response time
Avg The accumulated and average response time
Std Standard deviation of response time
Log2 Dist Geometric distribution based on logarithm 2
cmd_get Total get commands done
cmd_set Total set commands done
get_misses How many objects can’t be gotten from server
verify_misses How many objects need to verify but can’t get them
verify_failed How many objects with insistent value
expired_get How many objects are expired but we get them
unexpired_unget How many objects are unexpired but we can’t get them
written_bytes Total written bytes
read_bytes Total read bytes
object_bytes Total object bytes
packet_disorder How many UDP packages are disorder
packet_drop How many UDP packages are lost
udp_timeout How many times UDP time out happen
Run time Total run time
Ops Total operations
TPS Throughput, operations/second
Net_rate 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 tests, when connection is closed it will be reconnected.
-U, --udp
UDP tests, default memaslap 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.
memaslap -s 127.0.0.1:11211 -S 5s
memaslap -s 127.0.0.1:11211 -t 2m -v 0.2 -e 0.05 -b
memaslap -s 127.0.0.1:11211 -F config -t 2m -w 40k -S 20s -o 0.2
memaslap -s 127.0.0.1:11211 -F config -t 2m -T 4 -c 128 -d 20 -P 40k
memaslap -s 127.0.0.1:11211 -F config -t 2m -d 50 -a -n 40
memaslap -s 127.0.0.1:11211,127.0.0.1:11212 -F config -t 2m
memaslap -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://libmemcached.org/
Mingqiang Zhuang <[email protected]> (Schooner Technolgy) Brian Aker, <[email protected]>
memcached(1) libmemcached(3)
Brian Aker
2011-2013, Brian Aker DataDifferential, http://datadifferential.com/