neper is a Linux networking performance tool.
- Support multithreads and multi-flows out of the box.
- Generate workload with epoll.
- Collect statistics in a more accurate way.
- Export statistics to CSV for easier consumption by other tools.
neper currently supports eight workloads:
tcp_rrgenerates request/response workload (similar to HTTP or RPC) over TCPtcp_crris similar to tcp_rr but establishes a fresh connection for each request/response pairtcp_streamgenerates bulk data transfer workload (similar to FTP orscp) over TCPudp_rrgenerates request/response workload (similar to HTTP or RPC) over UDPudp_streamgenerates bulk data transfer workload (similar to FTP orscp) over UDPpsp_streamgenerates bulk data transfer workload (similar to FTP orscp) over TCP-PSPpsp_rrgenerates request/response workload (similar to HTTP or RPC) over TCP-PSPpsp_crris similar to psp_rr but establishes a fresh connection for each request/response pair
neper as a small code base with clean coding style and structure, and is easy to extend with new workloads and/or new options. It can also be embedded as a library in a larger application to generate epoll-based workloads.
Disclaimer: This is not an official Google product.
neper was at one time internally known as gperfnet. This documentation may use both names interchangeably.
This file documents the usage of neper as a tool.
Table of contents:
[TOC]
Currently neper consists of eight binaries, tcp_rr and tcp_stream and
tcp_crr and udp_rr and udp_stream and psp_stream and psp_crr and
psp_rr, corresponding to eight workloads.
neper is intended to be used between two machines. On each machine, the
neper process (e.g. tcp_rr or tcp_stream) spawns T threads (workers),
creates F flows (e.g. TCP connections), and multiplexes the F flows evenly over
the T threads. Each thread has a epoll set to manage multiple flows. Each
flow keeps its own state to make sure the expected workload is generated.
For ease of explanation, we refer to the two machines as the server and the
client. The server is the process which binds to an endpoint, while the client
is the process which connects to the server. The ordering of bind() and
connect() invocations is insignificant, as neper ensures the two processes
are synchronized properly.
When launching the client process, we can specify the number of seconds to generate network traffic "-l, --test-length". After that duration, the client will notify the server to stop and exit. To run the test again, we have to restart both the client and the server. This is different from netperf, and hopefully should make individual tests more independent from each other.
Optionally, a negative value for "-l, --test-length", a la netperf, can be specified where the absolute value is used for rr test to specify the number of transactions instead of time. The transacions are distributed across flows on demand. Note, the "-s, --suicide-length" option can be used to specify a time limit. Also, the negative value isn't currently supported for all tests, e.g. stream test, support for tests other than rr maybe added in the future.
Let's start the server process first:
server$ tcp_rr
percentiles=
all_samples=
port=12867
control_port=12866
host=
max_pacing_rate=0
interval=1.000000
pin_cpu=0
dry_run=0
client=0
buffer_size=65536
response_size=1
request_size=1
test_length=10
num_threads=1
num_flows=1
help=0
total_run_time=10
*(process waiting here)*
Immediately after the process starts, it prints several key=value pairs to
stdout. They are the command-line option values perceived by neper. In
this case, they are all default values. We can use them to verify the options
are parsed correctly, or to reproduce the test procedure from persisted
outputs.
At this point, the server is waiting for a client to connect. We can continue by running
client$ tcp_rr -c -H server
*(options key-value pairs omitted)*
-c is short for --client which means "this is the client process". If -c
is not specified, it will be started as a server and call bind(). When -c
is specified, it will try to connect(), and -H (short for --host)
specifies the server hostname to connect to. We can also use IP address
directly, to avoid resolving hostnames (e.g. through DNS).
For both bind() and connect(), we actually need the port number as well.
In the case of neper, two ports are being used, one for control plane, the
other one for data plane. Default ports are 12866 for control plane and 12867
for data plane. They can be overridden by -C (short for --control-port)
and -P (short for --port), respectively. Default port numbers are chosen
so that they don't collide with the port 12865 used by netperf. Note that for
UDP there is one data port per flow, allocated in a sequential range on the
server.
Immediately after the client process prints the options, it will connect to the server and start sending packets. After a period of time (by default 10 seconds), both processes print statistics summary to stdout and then exit:
server$ tcp_rr
*(previous output omitted)*
invalid_secret_count=0
time_start=1306173.666348235
utime_start=0.062141
utime_end=0.348902
stime_start=0.003883
stime_end=5.798208
maxrss_start=7896
maxrss_end=7896
minflt_start=568
minflt_end=571
majflt_start=7
majflt_end=7
nvcsw_start=26
nvcsw_end=329455
nivcsw_start=46
nivcsw_end=1028
num_transactions=329605
start_index=0
end_index=9
num_samples=10
throughput=33009.84
correlation_coefficient=1.00
time_end=1306183.666374314
client$ tcp_rr -c -H server
*(previous output omitted)*
*(new output lines are similar to the server)*
From the line throughput=33009.84, we know this test run finished 33009.84
request/response "ping-pong" transactions per second. A transaction for the
client means sending a request and then receiving a response. A transaction
for the server means receiving a request and then sending a response. The
number in this example is very high because it was run on localhost.
To look closer, let's reexamine the test parameters (command-line options), most importantly:
response_size=1
request_size=1
test_length=10
num_threads=1
num_flows=1
That means we were using one thread (on each side) with one flow (TCP connection between server and client) to send one-byte requests and responses over 10 seconds.
To run the test with 10 flows and two threads, we can instead use
server$ tcp_rr -F 10 -T 2
client$ tcp_rr -c -H server -F 10 -T 2
where -F is short for --num-flows and -T is short for --num-threads.
That will be 10 flows multiplexed on top of two threads, so normally it's 5
flows per thread. neper uses SO_REUSEPORT to load balance among the
threads, so it might not be exactly 5 flows per thread (e.g. may be 4 + 6).
This behavior might change in the future.
Server and client do not need to use the same number of threads. For example, we can create 2 threads on the server to serve requests from 4 threads from the client.
server$ tcp_rr -F 10 -T 2
client$ tcp_rr -c -H server -F 10 -T 4
In this case, the four client-side threads may handle 3 + 3 + 2 + 2 (= 10) flows respectively.
Also note that we have to specify the number of flows on the server side. This behavior might change in the future.
psp_crr and psp_rr shares all the command-line options and output with
tcp_crr and tcp_rr. They differ in the secure socket setup and the ensuing
encryption/decryption operations performed by the PSP-capable hardwares. More
details on PSP are available in: google/psp.
tcp_stream shares most of the command-line options with tcp_rr. They
differ in the output since for a bulk data transfer test like tcp_stream, we
care about the throughput in Mbps rather than in number of transactions.
By default, it's the client sending data to the server. We can enable the
other direction of data transfer (from server to client) by specifying
command-line options -r (short for --enable-read) and -w (short for
--enable-write).
server$ tcp_stream -w
client$ tcp_stream -c -H server -r
This is equivalent to
server$ tcp_stream -rw
client$ tcp_stream -c -H server -rw
since -w is auto-enabled for -c, and -r is auto-enabled when -c is
missing.
In both cases, the flows have bidirectional bulk data transfer. Previously,
netperf users may emulate this behavior with TCP_STREAM and TCP_MAERTS, at
the cost of doubling the number of netperf processes.
To have netperf TCP_MAERTS test, pass -M which overrides -r and -w.
server$ tcp_stream -M
client$ tcp_stream -c -H server -M
psp_stream shares all the command-line options and output with tcp_stream.
They differ in the secure socket setup and the ensuing encryption/decryption
operations performed by the PSP-capable hardwares. More details on PSP are
available in: google/psp.
client
host
local_hosts
control_port
port
mark
maxevents
num_flows
num_threads
test_length
pin_cpu
dry_run
logtostderr
nonblocking
buffer_size
all_samples
interval
max_pacing_rate
listen_backlog
tcp_tx_delay
request_size
response_size
percentiles
num_ports
request_size
response_size
percentiles
num_ports
request_size
response_size
percentiles
The output is only available in the detailed form (samples.csv) but not in
the stdout summary.
server$ ./tcp_rr
client$ ./tcp_rr -c -H server -A --percentiles=25,50,90,95,99
client$ cat samples.csv
time,tid,flow_id,bytes_read,transactions,latency_min,latency_mean,latency_max,latency_stddev,latency_p25,latency_p50,latency_p90,latency_p95,latency_p99,utime,stime,maxrss,minflt,majflt,nvcsw,nivcsw
2766296.649115114,0,0,31726,31726,0.000019,0.000030,0.008010,0.000086,0.000023,0.000026,0.000032,0.000033,0.000068,0.005268,0.479424,5288,71,0,28490,3360
2766297.649131797,0,0,62857,62857,0.000019,0.000031,0.007757,0.000078,0.000024,0.000027,0.000032,0.000034,0.000080,0.022667,0.933914,5288,133,0,57761,5692
2766298.649119440,0,0,98525,98525,0.000015,0.000027,0.004187,0.000048,0.000023,0.000025,0.000032,0.000033,0.000048,0.063623,1.481519,5288,204,0,91853,7383
2766299.649141269,0,0,138042,138042,0.000015,0.000024,0.009910,0.000091,0.000018,0.000018,0.000027,0.000030,0.000041,0.084147,1.984098,5288,283,0,129072,9754
2766300.649148147,0,0,169698,169698,0.000019,0.000030,0.004938,0.000063,0.000024,0.000027,0.000034,0.000036,0.000057,0.119381,2.493741,5288,346,0,160027,10551
2766301.649127545,0,0,202454,202454,0.000019,0.000029,0.006942,0.000060,0.000025,0.000027,0.000032,0.000032,0.000060,0.165496,2.920798,5288,411,0,186603,16817
2766302.649152705,0,0,234954,234954,0.000018,0.000029,0.012611,0.000100,0.000025,0.000026,0.000031,0.000032,0.000059,0.205488,3.349022,5288,475,0,212910,23195
2766303.649116145,0,0,269683,269683,0.000019,0.000027,0.004842,0.000038,0.000024,0.000026,0.000031,0.000032,0.000048,0.242531,3.806882,5288,544,0,240914,30076
2766304.649131298,0,0,302011,302011,0.000019,0.000030,0.004476,0.000049,0.000025,0.000029,0.000032,0.000033,0.000044,0.253141,4.294832,5288,608,0,270468,32944
2766305.649132278,0,0,340838,340838,0.000015,0.000025,0.000220,0.000006,0.000022,0.000025,0.000031,0.000033,0.000035,0.284624,4.808422,5288,685,0,308307,34005
reuseaddr
enable_read
enable_write
epoll_trigger
delay
throughput_opt
zerocopy
reuseaddr
delay
throughput_opt
zerocopy
reuseaddr
enable_read
enable_write
epoll_trigger
delay
throughput_opt
zerocopy
When consuming the key-value pairs in the output, the order of the keys should be insignificant. However, the keys are case sensitive.
total_run_time # expected time to finish, useful when combined with --dry-run
invalid_secret_count
time_start
start_index
end_index
num_samples
time_end
rusage
utime_start
utime_end
stime_start
stime_end
maxrss_start
maxrss_end
minflt_start
minflt_end
majflt_start
majflt_end
nvcsw_start
nvcsw_end
nivcsw_start
nivcsw_end
num_transactions
num_ports
throughput
correlation_coefficient # for throughput
num_transactions
num_ports
throughput
correlation_coefficient # for throughput
num_transactions
throughput_Mbps
correlation_coefficient # for throughput
throughput_units
throughput
num_transactions
throughput
correlation_coefficient # for throughput
num_transactions
throughput_Mbps
correlation_coefficient # for throughput
throughput_units
throughput
num_transactions
throughput_Mbps
correlation_coefficient # for throughput
throughput_units
throughput
- C99, avoid compiler-specific extensions.
- No external dependency.
- Linux coding style, with tabs expanded to 8 spaces.
- Breaking: changed histogram implementation