seqtest -- test TCP under load
seqtest is a program that is designed to verify correct end-to-end TCP
operation. It can also do some limited performance measurement. The biggest
advantage it has over alternative programs is that it allows one to alter
the traffic patterns and inject an element of randomness. This can be used
to help uncover race conditions and other difficult-to-reproduce bugs.
When run with a fixed number of iterations (see the count option below), and
with replies, it will report the round trip latency between each message sent,
and when its reply is received.
It runs in three modes:
* sender mode. In this mode seqtest sends a stream of packets, at some
(randomized, possibly zero) interval to the replier. It also runs a thread
to wait for replies from the replier, and verifies that messages are
delivered in order. Most of the customization is done when running the
sender side.
* synchoronous sender mode. THis mode is just like sender mode, except that
each thread waits for the reply before sending the next message. This is
a better measurement of latency.
* replier mode. In this mode, seqtest accepts incoming TCP connections,
analyzes them for correctness (in order delivery, etc.) and optionally
sends a reply. The size of the reply can vary, as well as the frequency
with which replies are made. A random delay can also be inserted before
sending the reply.
In sender mode, the synopsis is as follows:
seqtest -s [-d] [-o <option>=<value>[,<option>=<value>...] <address>...
Synchronous sender juse replaces -s for -S:
seqtest -S [-d] [-o <option>=<value>[,<option>=<value>...] <address>...
Where <address> has one of two forms:
<remote_addr>:<remote_port>
<local_addr>,<remote_addr>:<remote_port> (bind to local_addr)
The options are name value pairs; the following are defined:
ssize=<num> The size of message payload to send. Will be
rounded up to 40 bytes if less than that is specified,
as seqtest needs 40 bytes of header information on
each message. Currently messages are limited to 8000
bytes maximum, as well.
ssize_min=<num> A minimum value to use for send payload sizes. If
this is specified, then each sent message will have a
random payload size of at least this size.
ssize_max=<num> A maximum value to use for send paylod size.
rsize=<num> The size of reply payloads to send. As with ssize,
the value must be between 40 and 8000, inclusive.
rsize_min=<num> A minimum reply payload size, used when randomly
choosing reply payload sizes. Each reply's size
is generated randomly.
rsize_max=<num> A maximum reply payload sized, used when randomly
choosing reply payload sizes.
threads=<num> The number of parallel sending threads to spawn.
There actually be twice this number of threads
spawned. One for each connection is spawned for
sending messages and another for receiving replies.
This corresponds to the number of concurrent TCP
flows to process.
sdelay=<ns> A number of nanoseconds (can be zero) to wait
between sending messages. If the number is at least
1000000 then a sleeping delay will be used, otherwise
the sender will spin on the CPU.
sdelay_min=<ns> Allows the delay before sending a message to be
randomized each time. This is a minimum number
of nsec.
sdelay_max=<ns> Maximum time to wait before sending.
rdelay=<ns> As with sdelay, this is a wait time, but it is the
time that a replier should wait before sending
a reply. This can be randomized for each
reply with rdelay_min and rdelay_max. This simulates
work being done before sending a reply.
rinterval=<num> The interval between replies, as a number of messages
received. For example, if 2, then a reply will only
be sent every other message. Defaults to 1. If 0,
then no replies are sent at all.
count=<num> The number of messages each sending thread should send.
The address(es) are IP address (or hostname) and port pairs separated by
a colon to use for connecting. If a name resolves to multiple IP addresses,
then multiple senders will be spawned by default, one for each resolved IP.
(The exception would be if a value of threads is specified in above. In
that case exactly that many senders would be spawned, balancing the load
amongst resolved addreses.) For example, to send traffic to localhost port 88,
use 127.0.0.1:88 or "localhost:88". Note that the addresses can be either IPv4
or IPv6.
The receiver synopsis is simpler:
seqtest -r <address>...
In this case the list of addresses is specified just as with the sender, but
the addresses must be addresses local to the host where seqtest is run. The
program will bind and listen for incoming connections on these addresses,
and reply according to the specifications of received messages.