twemperf (pronounced "two-em-perf"), aka mcperf is a tool for measuring memcached server performance. mcperf is like httperf, but for memcached protocol. It speaks memcached ASCII protocol and is capable of generating connections and requests at a high rate.
To build mcperf from distribution tarball:
$ ./configure
$ make
$ sudo make install
To build mcperf from distribution tarball in debug mode:
$ CFLAGS="-ggdb3 -O0" ./configure --enable-debug
$ make
$ sudo make install
To build mcperf from source in debug mode:
$ git clone git://github.com/twitter/twemperf.git
$ cd twemperf
$ autoreconf -fvi
$ CFLAGS="-ggdb3 -O0" ./configure --enable-debug
$ make
$ src/mcperf -h
Usage: mcperf [-?hV] [-v verbosity level] [-o output file]
[-s server] [-p port] [-H] [-t timeout] [-l linger]
[-b send-buffer] [-B recv-buffer] [-D]
[-m method] [-e expiry] [-q] [-P prefix]
[-c client] [-n num-conns] [-N num-calls]
[-r conn-rate] [-R call-rate] [-z sizes]
Options:
-h, --help : this help
-V, --version : show version and exit
-v, --verbosity=N : set logging level (default: 5, min: 0, max: 11)
-o, --output=S : set logging file (default: stderr)
-s, --server=S : set the hostname of the server (default: localhost)
-p, --port=N : set the port number of the server (default: 11211)
-H, --print-histogram : print response time histogram
...
-t, --timeout=X : set the connection and response timeout in sec (default: 0.0 sec)
-l, --linger=N : set the linger timeout in sec, when closing TCP connections (default: off)
-b, --send-buffer=N : set socket send buffer size (default: 4096 bytes)
-B, --recv-buffer=N : set socket recv buffer size (default: 16384 bytes)
-D, --disable-nodelay : disable tcp nodelay
...
-m, --method=M : set the method to use when issuing memcached request (default: set)
-e, --expiry=N : set the expiry value in sec for generated requests (default: 0 sec)
-q, --use-noreply : set noreply for generated requests
-P, --prefix=S : set the prefix of generated keys (default: mcp:)
...
-c, --client=I/N : set mcperf instance to be I out of total N instances (default: 0/1)
-n, --num-conns=N : set the number of connections to create (default: 1)
-N, --num-calls=N : set the number of calls to create on each connection (default: 1)
-r, --conn-rate=R : set the connection creation rate (default: 0 conns/sec)
-R, --call-rate=R : set the call creation rate (default: 0 calls/sec)
-z, --sizes=R : set the distribution for item sizes (default: d1 bytes)
...
Where:
N is an integer
X is a real
S is a string
M is a method string and is either a 'get', 'gets', 'delete', 'cas', 'set', 'add', 'replace'
'append', 'prepend', 'incr', 'decr'
R is the rate written as [D]R1[,R2] where:
D is the distribution type and is either deterministic 'd', uniform 'u', or exponential 'e' and if:
D is ommited or set to 'd', a deterministic interval specified by parameter R1 is used
D is set to 'e', an exponential distibution with mean interval of R1 is used
D is set to 'u', a uniform distribution over interval [R1, R2) is used
R is 0, the next request or connection is created after the previous one completes
- Single threaded.
- Asynchronous I/O through non-blocking sockets and Linux epoll(7) syscall.
- Horizonal scaling through many concurrent mcperf processes on several different machines.
mcperf is composed of three main subsystems:
- Core Engine
- Load Generator
- Stats Collector
The core engine is responsible for the event handling and parsing of memcached protocol and drives the main event loop.
The load generator is resposible for generating load either periodically or as a one-shot event. Each load generator is implemented as a self contained module. Some examples of load generators are:
- Connection generator creates connections to a server at a given rate.
- Call generator issue calls (requests) on a connection at a given rate.
- Size generator generates item sizes.
The stats collector is responsible for collecting statistics. Each stats collector is implemented as a self contained module. Some examples of stats collectors are:
- Connection stats collects connection stats.
- Call stats collects call (request and response) stats.
The following example creates 1000 connections to a memcached server running on localhost:11211. The connections are created at the rate of 1000 conns/sec and on every connection it sends 10 'set' requests at the rate of 1000 reqs/sec with the item sizes derived from a uniform distribution in the interval of [1,16) bytes.
$ mcperf --linger=0 --timeout=5 --conn-rate=1000 --call-rate=1000 --num-calls=10 --num-conns=1000 --sizes=u1,16
Total: connections 1000 requests 10000 responses 10000 test-duration 1.009 s
Connection rate: 991.1 conn/s (1.0 ms/conn <= 23 concurrent connections)
Connection time [ms]: avg 10.3 min 10.1 max 14.1 stddev 0.1
Connect time [ms]: avg 0.2 min 0.1 max 0.8 stddev 0.0
Request rate: 9910.5 req/s (0.1 ms/req)
Request size [B]: avg 35.9 min 28.0 max 44.0 stddev 4.8
Response rate: 9910.5 rsp/s (0.1 ms/rsp)
Response size [B]: avg 8.0 min 8.0 max 8.0 stddev 0.0
Response time [ms]: avg 0.2 min 0.1 max 13.4 stddev 0.00
Response time [ms]: p25 1.0 p50 1.0 p75 1.0
Response time [ms]: p95 1.0 p99 1.0 p999 1.0
Response type: stored 10000 not_stored 0 exists 0 not_found 0
Response type: num 0 deleted 0 end 0 value 0
Response type: error 0 client_error 0 server_error 0
Errors: total 0 client-timo 0 socket-timo 0 connrefused 0 connreset 0
Errors: fd-unavail 0 ftab-full 0 addrunavail 0 other 0
CPU time [s]: user 0.64 system 0.35 (user 63.6% system 35.1% total 98.7%)
Net I/O: bytes 428.7 KB rate 424.8 KB/s (3.5*10^6 bps)
The following example creates 100 connections to a memcached server running on localhost:11211. Every connection is created after the previous connection is closed. On every connection we send 100 'set' requests and every request is created after we have received the response for the previous request. All the set requests generated have a fixed item size of 1 byte.
$ mcperf --linger=0 --call-rate=0 --num-calls=100 --conn-rate=0 --num-conns=100 --sizes=d1
Total: connections 100 requests 10000 responses 10000 test-duration 1.268 s
Connection rate: 78.9 conn/s (12.7 ms/conn <= 1 concurrent connections)
Connection time [ms]: avg 12.7 min 12.6 max 13.5 stddev 0.1
Connect time [ms]: avg 0.0 min 0.0 max 0.1 stddev 0.0
Request rate: 7886.1 req/s (0.1 ms/req)
Request size [B]: avg 28.0 min 28.0 max 28.0 stddev 0.0
Response rate: 7886.1 rsp/s (0.1 ms/rsp)
Response size [B]: avg 8.0 min 8.0 max 8.0 stddev 0.0
Response time [ms]: avg 0.1 min 0.1 max 1.0 stddev 0.00
Response time [ms]: p25 1.0 p50 1.0 p75 1.0
Response time [ms]: p95 1.0 p99 1.0 p999 1.0
Response type: stored 10000 not_stored 0 exists 0 not_found 0
Response type: num 0 deleted 0 end 0 value 0
Response type: error 0 client_error 0 server_error 0
Errors: total 0 client-timo 0 socket-timo 0 connrefused 0 connreset 0
Errors: fd-unavail 0 ftab-full 0 addrunavail 0 other 0
CPU time [s]: user 0.51 system 0.75 (user 40.0% system 59.0% total 99.0%)
Net I/O: bytes 351.6 KB rate 277.2 KB/s (2.3*10^6 bps)
Have a bug or a question? Please create an issue here on GitHub!
https://github.com/twitter/twemperf/issues
- Manju Rajashekhar (@manju)
- Clojure Janet (@clojurejanet)