Memcache-perf

Memcache-perf is a memcached load generator designed for high request rates, good tail-latency measurements, and realistic request stream generation.

Requirements

1. A C++0x compiler
2. libevent2 (get headers and install build-dev for memcached for rest)
3. zeromq

Tested on ubuntu 14.04,16.04,18.04 x86 64b and ARMv8.

Building

apt-get install libevent-dev libzmq3-dev
apt-get build-dep memcached
make

Basic Usage

Type './mcperf -h' for a full list of command-line options. At minimum, a server must be specified.

$ ./mcperf -s localhost
#type       avg     std     min      p5     p10     p50     p67     p75     p80     p85     p90     p95     p99    p999   p9999
read       62.2    24.1    54.7    59.1    59.4    61.5    62.3    62.8    63.0    63.3    63.5    63.8    68.8    80.2  1012.5
update      0.0     0.0     0.0     0.0     0.0     0.0     0.0     0.0     0.0     0.0     0.0     0.0     0.0     0.0     0.0
op_q        1.0     0.0     1.0     1.0     1.0     1.0     1.1     1.1     1.1     1.1     1.1     1.1     1.1     1.1     1.1

Total QPS = 16082.2 (80411 / 5.0s)

Misses = 0 (0.0%)
Skipped TXs = 0 (0.0%)

RX   19861517 bytes :    3.8 MB/s
TX    2894832 bytes :    0.6 MB/s
CPU Usage Stats (avg/min/max): 2.21%,0.38%,4.05%

mcperf reports the latency (average, minimum, and various percentiles) for get and set commands, as well as achieved QPS and network goodput. A separate thread is also keeping track of client CPU usage on the master. A warning will be issued if the master client CPU usage goes above 95%. In that case, it is recommended to add more machines as agents. If verbose (-v) flag is enabled on the an agent, it will report it's cpu usage as well.

To achieve high request rate, you must configure mcperf to use multiple threads, multiple connections, connection pipelining, or remote agents.

$ ./mcperf -s zephyr2-10g -T 24 -c 8
#type       avg     min     1st     5th    10th    90th    95th    99th
read      598.8    86.0   437.2   466.6   482.6   977.0  1075.8  1170.6
update      0.0     0.0     0.0     0.0     0.0     0.0     0.0     0.0
op_q        1.5     1.0     1.0     1.1     1.1     1.9     1.9     2.0

Total QPS = 318710.8 (1593559 / 5.0s)

Misses = 0 (0.0%)

RX  393609073 bytes :   75.1 MB/s
TX   57374136 bytes :   10.9 MB/s

Suggested Usage

Real deployments of memcached often handle the requests of dozens, hundreds, or thousands of front-end clients simultaneously. However, by default, mcperf establishes one connection per server and meters requests one at a time (it waits for a reply before sending the next request). This artificially limits throughput (i.e. queries per second), as the round-trip network latency is almost certainly far longer than the time it takes for the memcached server to process one request.

In order to get reasonable benchmark results with mcperf, it needs to be configured to more accurately portray a realistic client workload. In general, this means ensuring that (1) there are a large number of client connections, (2) there is the potential for a large number of outstanding requests, and (3) the memcached server saturates and experiences queuing delay far before mcperf does. I suggest the following guidelines:

1. Establish more than 50 connections per memcached server thread.
2. Don't exceed more than about 10 connections per mcperf thread.
3. Use multiple mcperf agents in order to achieve (1) and (2).
4. Do not use more mcperf threads than hardware cores/threads.

Here's an example:

memcached_server$ memcached -t 4 -c 32768
agent1$ mcperf -T 16 -A
agent2$ mcperf -T 16 -A
agent3$ mcperf -T 16 -A
agent4$ mcperf -T 16 -A
agent5$ mcperf -T 16 -A
agent6$ mcperf -T 16 -A
agent7$ mcperf -T 16 -A
agent8$ mcperf -T 16 -A
master$ mcperf -s memcached_server --loadonly
master$ mcperf -s memcached_server --noload \
    -B -T 16 -Q 1000 -D 4 -C 4 \
    -a agent1 -a agent2 -a agent3 -a agent4 \
    -a agent5 -a agent6 -a agent7 -a agent8 \
    -c 4 -q 200000

This will create 8*16*4 = 512 connections total, which is about 128 per memcached server thread. This ought to be enough outstanding requests to cause server-side queuing delay, and no possibility of client-side queuing delay adulterating the latency measurements.

Command-line Options

mcperf 0.4

Usage: mcperf -s server[@port] [options]

"High-performance" memcached benchmarking tool

-h, --help Print help and exit --version Print version and exit -v, --verbose Verbosity. Repeat for more verbose. --quiet Disable log messages.

Basic options: -s, --server=STRING Memcached server hostname[:port[-end_port]]. Repeat to specify multiple servers. --binary Use binary memcached protocol instead of ASCII. -q, --qps=INT Target aggregate QPS. 0 = peak QPS. (default=0') -t, --time=INT Maximum time to run (seconds). (default=5') --profile=INT Select one of several predefined profiles. -K, --keysize=STRING Length of memcached keys (distribution). (default=30') --keyorder=STRING Selection of memcached keys to use (distribution). (default=none') -V, --valuesize=STRING Length of memcached values (distribution). (default=200') -r, --records=INT Number of memcached records to use. If multiple memcached servers are given, this number is divided by the number of servers. (default=10000') -u, --update=FLOAT Ratio of set:get commands. (default=`0.0')

Advanced options: -U, --username=STRING Username to use for SASL authentication. -P, --password=STRING Password to use for SASL authentication. -T, --threads=INT Number of threads to spawn. (default=1') --affinity Set CPU affinity for threads, round-robin -c, --connections=INT Connections to establish per server. (default=1') -d, --depth=INT Maximum depth to pipeline requests. (default=1') -R, --roundrobin Assign threads to servers in round-robin fashion. By default, each thread connects to every server. -i, --iadist=STRING Inter-arrival distribution (distribution). Note: The distribution will automatically be adjusted to match the QPS given by --qps. (default=exponential') -S, --skip Skip transmissions if previous requests are late. This harms the long-term QPS average, but reduces spikes in QPS after long latency requests. --moderate Enforce a minimum delay of ~1/lambda between requests. --noload Skip database loading. --loadonly Load database and then exit. -B, --blocking Use blocking epoll(). May increase latency. --no_nodelay Don't use TCP_NODELAY. -w, --warmup=INT Warmup time before starting measurement. -W, --wait=INT Time to wait after startup to start measurement. --save=STRING Record latency samples to given file. --search=N:X 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). --scan=min:max:step Scan latency across QPS rates from min to max. -e, --trace To enable server tracing based on client activity, will issue special start_trace/stop_trace commands. Requires memcached to support these commands. -G, --getq_size=INT Size of queue for multiget requests. (default=100') -g, --getq_freq=FLOAT Frequency of multiget requests, 0 for no multi-get, 100 for only multi-get. (default=0.0') --keycache_capacity=INT Cached key capacity. (default 10000) (default=10000') --keycache_reuse=INT Number of times to reuse key cache before generating new req sequence. (Default 100) (default=100') --keycache_regen=INT When regenerating control number of requests to regenerate. (Default 1%) (default=`1') --plot_all Create plot/csv of latency histogram at each step when using gnuplot and loghistogram sampler

Agent-mode options: -A, --agentmode Run client in agent mode. -a, --agent=host[@agent_port] Enlist remote agent, with optional agent port to support multiple agents per host. -p, --agent_port=STRING Agent port, ignored when it is part of the agent option (default=5556') -l, --lambda_mul=INT Lambda multiplier. Increases share of QPS for this client. (default=1') -C, --measure_connections=INT Master client connections per server, overrides --connections. -Q, --measure_qps=INT Explicitly set master client QPS, spread across threads and connections. -D, --measure_depth=INT Set master client connection depth. -m, --poll_freq=INT Set frequency in seconds for agent protocol recv polling. (default=1') -M, --poll_max=INT Set timeout for agent protocol recv polling. An agent not responding within time limit will be dropped. (default=120')

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.

Predefined profiles to approximate some use cases:

1. memcached for web serving benchmark : p95, 20ms, FB key/value/IA, >4000 connections to the device under test.
2. memcached for applications backends : p99, 10ms, 32B key , 1000B value, uniform IA, >1000 connections
3. memcached for low latency (e.g. stock trading): p99.9, 32B key, 200B value, uniform IA, QPS rate set to 100000
4. P99.9, 1 msec. Key size = 32 bytes; value size has uniform distribution from 100 bytes to 1k;
5. Update heavy large queue memcached: limit connections and increase depth to 148, update ratio 0.623595
6. Update heavy high latency memcached: Single connection and increase depth to 249, update ratio 0.558257
7. Small DB low connection memcached: 1000 records, depth 207, 6 connections per thread
8. Small DB low connection large queue memcached: 1000 records, depth 250, 7 connections per thread
9. Small DB low connection large queue with updates memcached: 1000 records, depth 250, 7 connections per thread, update ratio 0.001114

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

[fixed:] Always generates . uniform: Uniform distribution between 0 and . normal:, Normal distribution. exponential: Exponential distribution. pareto:,, Generalized Pareto distribution. gev:,, 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