package "mutilate3"
version "0.1"
usage "mutilate -s server[:port] [options]"
description "\"High-performance\" memcached benchmarking tool"

args "-c cc --show-required -C --default-optional -l"

option "verbose" v "Verbosity. Repeat for more verbose." multiple
option "quiet" - "Disable log messages."

text "\nBasic options:"

option "server" s "Memcached server hostname[:port].  \
Repeat to specify multiple servers." string multiple
option "binary" - "Use binary memcached protocol instead of ASCII."
option "qps" q "Target aggregate QPS. 0 = peak QPS." int default="0"
option "time" t "Maximum time to run (seconds)." int default="5"

option "keysize" K "Length of memcached keys (distribution)."
       string default="30"
option "valuesize" V "Length of memcached values (distribution)."
       string default="200"

option "records" r "Number of memcached records to use.  \
If multiple memcached servers are given, this number is divided \
by the number of servers." int default="10000"

option "update" u "Ratio of set:get commands." float default="0.0"

text "\nAdvanced options:"

option "username" U "Username to use for SASL authentication." string
option "password" P "Password to use for SASL authentication." string
option "threads" T "Number of threads to spawn." int default="1"
option "connections" c "Connections to establish per server." int default="1"
option "depth" d "Maximum depth to pipeline requests." int default="1"
option "roundrobin" R "Assign threads to servers in round-robin fashion.  \
By default, each thread connects to every server."

option "iadist" i "Inter-arrival distribution (distribution).  Note: \
The distribution will automatically be adjusted to match the QPS given \
by --qps." string default="exponential"
option "skip" S "Skip transmissions if previous requests are late.  This \
harms the long-term QPS average, but reduces spikes in QPS after \
long latency requests."

option "noload" - "Skip database loading."
option "loadonly" - "Load database and then exit."

option "blocking" B "Use blocking epoll().  May increase latency."
option "no_nodelay" - "Don't use TCP_NODELAY."

option "warmup" w "Warmup time before starting measurement." int
option "wait" W "Time to wait after startup to start measurement." int

option "search" - "Search for the QPS where N-order statistic < Xus.  \
(i.e. --search 95:1000 means find the QPS where 95% of requests are \
faster than 1000us)." string typestr="N:X"
option "scan" - "Scan latency across QPS rates from min to max."
       string typestr="min:max:step"

text "\nAgent-mode options:"
option "agentmode" A "Run client in agent mode."
option "agent" a "Enlist remote agent." string typestr="host" multiple
option "lambda_mul" l "Lambda multiplier.  Increases share of \
QPS for this client." int default="1"
option "measure_connections" C "Master client connections per server, \
overrides --connections." int
option "measure_qps" Q "Explicitly set master client QPS, \
spread across threads and connections." int
option "measure_depth" D "Set master client connection depth." int

text "
The --measure_* options aid in taking latency measurements of the
memcached server without incurring significant client-side queuing
delay.  --measure_connections allows the master to override the
--connections option.  --measure_depth allows the master to operate as
an \"open-loop\" client while other agents continue as a regular
closed-loop clients.  --measure_qps lets you modulate the QPS the
master queries at independent of other clients.  This theoretically
normalizes the baseline queuing delay you expect to see across a wide
range of --qps values.
"

text "
Some options take a 'distribution' as an argument.
Distributions are specified by <distribution>[:<param1>[,...]].
Parameters are not required.  The following distributions are supported:

   [fixed:]<value>              Always generates <value>.
   uniform:<max>                Uniform distribution between 0 and <max>.
   normal:<mean>,<sd>           Normal distribution.
   exponential:<lambda>         Exponential distribution.
   pareto:<loc>,<scale>,<shape> Generalized Pareto distribution.
   gev:<loc>,<scale>,<shape>    Generalized Extreme Value distribution.

   To recreate the Facebook \"ETC\" request stream from [1], the
   following hard-coded distributions are also provided:

   fb_value   = a hard-coded discrete and GPareto PDF of value sizes
   fb_key     = \"gev:30.7984,8.20449,0.078688\", key-size distribution
   fb_ia      = \"pareto:0.0,16.0292,0.154971\", inter-arrival time dist.

[1] Berk Atikoglu et al., Workload Analysis of a Large-Scale Key-Value Store,
    SIGMETRICS 2012
"