httperf.man

一个WEB服务器的性能测试工具

Thus, if the URI-prefix is
.BR /wset1024 ,
then the URI being accessed would be
.BR /wset1024/0/1/0/3.html .
In other words, the files on the server need to be organized as a
10ary tree.
.SH OUTPUT
This section describes the statistics output at the end of each test
run.  The basic information shown below is printed independent of the
selected workload generator.
.PP
.RS
.br
.B Total:
connections 30000 requests 29997 replies 29997 test-duration 299.992 s
.PP
.B Connection rate:
100.0 conn/s (10.0 ms/conn, <=14 concurrent connections)
.br
.B Connection time [ms]:
min 1.4 avg 3.0 max 163.4 median 1.5 stddev 7.3
.br
.B Connection time [ms]:
connect 0.6
.br
.B Connection length [replies/conn]:
1.000
.PP
.B Request rate:
100.0 req/s (10.0 ms/req)
.br
.B Request size [B]:
75.0
.PP
.B Reply rate [replies/s]:
min 98.8 avg 100.0 max 101.2 stddev 0.3 (60 samples)
.br
.B Reply time [ms]:
response 2.4 transfer 0.0
.br
.B Reply size [B]:
header 242.0 content 1010.0 footer 0.0 (total 1252.0)
.br
.B Reply status:
1xx=0 2xx=29997 3xx=0 4xx=0 5xx=0
.PP
.B CPU time [s]:
user 94.31 system 205.26 (user 31.4% system 68.4% total 99.9%)
.br
.B Net I/O:
129.6 KB/s (1.1*10^6 bps)
.PP
.B Errors:
total 3 client-timo 0 socket-timo 0 connrefused 3 connreset 0
.br
.B Errors:
fd-unavail 0 addrunavail 0 ftab-full 0 other 0
.br
.RE
.PP
There are six groups of statistics: overall results (``Total''),
connection related results (``Connection''), results relating to the
issuing of HTTP requests (``Request''), results relating to the replies
received from the server (``Reply''), miscellaneous results relating to
the CPU (``CPU'') and network (``Net I/O'') utilization and, last but not
least, a summary of errors encountered (``Errors'').
.TP
Total Section
.br
This section summarizes how many TCP connections were initiated by
.BR httperf ,
how many requests it sent out, how many replies it received, and
what the total test duration was.  In the example output
shown above, 30,000 connections were created, 29,997 requests were
sent out and 29,997 replies were received.  The duration of the
test was almost exactly 5 minutes (300 seconds).
.TP
Connection Section
.br
This section conveys information related to TCP connections generated
by the tool.  Specifically, the ``Connection rate'' line shows that new
connections were initiated at a rate of 100.0 connections per second.
This rate corresponds to a period of 10.0 milliseconds per
connection.  The last number in this line shows that at most 14
connections were open at any given time.

The first line labeled ``Connection time'' gives lifetime statistics
for successful connections.  The lifetime of a connection is the time
between a TCP connection is initiated and the time the connection is
closed.  A connection is considered successful if it had at least one
call that completed successfully.  In the example output, the line
indicates that the minimum (``min'') connection lifetime was 1.4
milliseconds, the average (``avg'') lifetime was 3.0 milliseconds, the
maximum (``max'') was 163.4 milliseconds, the median (``median'')
lifetime was 1.5 milliseconds, and that the standard deviation of the
lifetimes was 7.3 milliseconds.  The median lifetime is computed based
on a histogram with one millisecond resolution and a maximum lifetime
of 100 seconds.  Thus, the median is accurate to within half a
millisecond if at least half of the successful connections have a
lifetime of no more than 100 seconds.

The next statistic in this section is the average time it took to
establish a TCP connection.  Only successful TCP connection
establishments are counted.  In the example, the second line labeled
``Connection time'' shows that, on average, it took 0.6 milliseconds
to establish a connection.

The final line in this section is labeled ``Connection length.''  It
gives the average number of replies received on each connection that
received at least one reply (i.e., connections that failed before
yielding the first reply are not counted).  This number can be bigger
than 1.0 due to persistent connections.
.TP
Request Section
.br
The line labeled ``Request rate'' gives the rate at which HTTP requests
were issued and the period that this rate corresponds to.  In the
example above, the request rate was 100.0 requests per second, which
corresponds to 10.0 milliseconds per request.  As long as no
persistent connections are employed, the results in this section are
very similar or identical to results in the connection section.
However, when persistent connections are used, several calls can be
performed on a single connection in which case the results would be
different.

The line labeled ``Request size'' gives the average size of the HTTP
requests in bytes.  In the example above, the average request size was
75 bytes.
.TP
Reply Section
.br
For simple measurements, this section is often the most interesting
one as the line labeled ``Reply rate'' gives various statistics for
the reply rate.  In the example above, the minimum (``min'') reply
rate was 98.8 replies per second, the average (``avg'') was 100
replies per second, and the maximum (``max'') rate was 101.2 replies
per second.  The standard deviation was 0.3 replies per second.  The
number enclosed in parentheses shows that 60 reply rate samples were
acquired.  At present,
.B httperf
collects a rate sample once every five seconds.  To obtain a
meaningful standard deviation, it is recommended to run tests long
enough so at least thirty samples are obtained.  This corresponds to a
test duration of at least 150 seconds.

The line labeled ``Reply Time'' gives information on how long it took
for the server to respond and how long it took to receive the reply.
In the example, it took on average 2.4 milliseconds between sending
the first byte of the request and receiving the first byte of the
reply.  The time to ``transfer'', or read, the reply was too short to
be measured, so it shows up as zero.  The is typical when the entire
reply fits into a single TCP segment.

The next line, labeled ``Reply size'' contains statistics on the
average size of the replies---all numbers are in reported bytes.
Specifically, the line lists the average length of reply headers, the
content, and footers (HTTP/1.1 uses footers to realize the ``chunked''
transfer encoding).  For convenience, the average total number of
bytes in the replies is also given in parentheses.  In the example,
the average header length (``header'') was 242 bytes, the average
content length (``content'') was 1010 bytes, and there were no footers
(``footer'' length is zero).  The total reply length of 1252 bytes on
average.

The final line in this section is a histogram of the major status
codes received in the replies from the server.  The major status code
is the ``hundreds''-digit of the full HTTP status code.  In the
example, all 29,997 replies had a major status code of 2.  It's a good
guess that all status codes were ``200 OK'' but the information in the
histogram is not detailed enough to allow distinguishing status codes
with the same major code.
.TP
Miscellaneous Section
.br
This section starts with a summary of the CPU utilization on the
client machine.  In the example, the line labeled ``CPU time'' shows
that 94.31 seconds were spent executing in user mode (``user''),
205.26 seconds were spent executing in system mode (``system'') and
that this corresponds to 31.4% user mode execution and 68.4% system
execution.  The total utilization was 99.9%, which is expected given
that
.B httperf
is a CPU hog.  A total CPU utilization of significantly less than 100%
is a sign that there were competing processes that interfered with the
test.

The line labeled ``Net I/O'' gives the average network throughput in
kilobytes per second (where a kilobyte is 1024 bytes) and in megabits
per second (where a megabit is 10^6 bits).  In the example, an average
network usage of about 129.6 kilobytes per second was sustained.  The
number in parentheses shows that this corresponds to about 1.1
megabits per second.  This network bandwidth is computed based on the
number of bytes sent and received on the TCP connections.  In other
words, it does not account for the network headers or TCP
retransmissions that may have occurred.
.TP
Errors Section
.br
The last section contains statistics on the errors that were
encountered during a test.  In the example, the two lines labeled
``Errors'' show that there were a total of three errors and that all
three errors were due to the server refusing to accept a connection
(``connrefused'').  A description of each error counter follows:

.B client-timo:
The number of times a session, connection, or call failed due
to a client timeout (as specified by the
.B --timeout
and
.BR --think-timeout )
options.

.B socket-timo:
The number of times a TCP connection failed with a socket-level
timeout (ETIMEDOUT).

.B connrefused:
The number of times a TCP connection attempt failed with a
``connection refused by server'' error (ECONNREFUSED).

.B connreset:
The number of times a TCP connection failed due to a RESET from the
server.  Typically, a RESET is received when the client attempts to
send data to the server at a time the server has already closed its
end of the connection.  NT servers also send RESETs when attempting to
establish a new connection when the listen queue is full.

.B fd-unavail:
The number of times the
.B httperf
process was out of file descriptors.  Whenever this count is non-zero,
the test results are meaningless because the client was overloaded
(see section "CHOOSING TIMEOUT VALUES").

.B addrunavail:
The number of times the client was out of TCP port numbers
(EADDRNOTAVAIL).  This error should never occur.  If it does, the
results should be discarded.

.B ftab-full:
The number of times the system's file descriptor table is full.
Again, this error should never occur.  If it does, the results should
be discarded.

.B other:
The number of times some other type of error occurred.  Whenever this
counter is non-zero, it is necessary to track down the real cause of
the error.  To assist in doing this,
.B httperf
prints the error code (errno) of the first unknown errors that occurs
during a test run.
.RE
.PP
When
.B --wsess
or
.B --wsesslog
is specified,
.B httperf
generates and measures sessions instead of individual calls and
additional statistics are printed at the end of a test.  An example
output is shown below.
.PP
.RS
.B Session rate [sess/s]:
min 0.00 avg 0.59 max 2.40 stddev 0.37 (240/450)
.br
.B Session:
avg 6.45 connections/session
.br
.B Session lifetime [s]:
123.9
.br
.B Session failtime [s]:
58.5
.br
.B Session length histogram:
4 7 4 ... 3 3 240
.RE
.PP
The line labeled ``Session rate'' shows the minium, average, and
maximum rate at which sessions completed (based on a 5 second sampling
interval).  It also shows the standard deviation of the session
completion rate.  The numbers in parentheses show how many sessions
succeeded and how many sessions were initiated.  In the example above,
the minimum, average, and maximum session completion rates were 0.00,
0.59, and 2.40 sessions per second, respectively.  The standard
deviation was 0.37 sessions per second and 240 out of 450 sessions
completed successfully (210 failed due to errors such as timeouts).

The next line, labeled ``Session:'' shows the average length of a
session measured in connections.  In the example above, an average of
6.45 connections were required to complete a session.

The line labeled ``Session lifetime'' gives the average time it took
to complete a successful session.  In the example above, it took an
average of 123.9 seconds.

The line labeled ``Session failtime'' gives the average time it took
before an unsuccessful session failed.  In the example above, it took
on average 58.5 seconds for a session to fail.

Finally, the line labeled ``Session length histogram'' gives a
histogram of the number of replies received by each session.  In the
example above, 4 sessions ended after receiving no reply at all, 7
ended after receiving one reply, and so on (the ellipsis indicates
additional histogram counts that were omitted from this manual for
space reasons).  Note that this histogram does not distinguish between
successful and failed sessions.

.SH CHOOSING TIMEOUT VALUES
Since the machine that
.B httperf
runs on has only a finite set of resource available, it can not
sustain arbitrarily high HTTP loads.  For example, one limiting factor
is that there are only roughly 60,000 TCP port numbers that can be in
use at any given time.  Since on most UNIX systems it takes one minute
for a TCP connection to be fully closed (leave the TIME_WAIT state),
the maximum rate a client can sustain is at most 1,000 requests per
second.

The actual sustainable rate is often much lower than that because
before running out of TCP ports, the machine is likely to run out of
file descriptors (one file descriptor is used up for each open TCP
connection).  By default, HP-UX 10.20 allows 1,024 open file
descriptors per process.  This means that without extra precautions,
.B httperf
could potentially very quickly use up all available file descriptors,
at which point it could not induce any additional load on the server.
To avoid this problem,
.B httperf
provides option
.B --timeout
to set a timeout for all communication with the server.  If the server
does not respond before the timeout expires, the client considers the
corresponding session, connection, or call to be ``dead,'' closes the
associated TCP connection, and increases the ``client-timo'' error
count.  The only exception to this rule is that after sending an
entire request to the server,
.B httperf
allows the server to take some additional time before it starts
sending the reply.  This is to accommodate HTTP requests that take a
long time to complete on the server.  This additional time is called
the ``server think time'' and can be specified by option
.BR --think-timeout .
By default, this additional think time is zero seconds, so the server
would always have to respond within the time alloted by option
.BR --timeout .

Timeouts allow
.B httperf 
to sustain high offered loads even when the server is overloaded.  For
example, with a timeout of 2 seconds and assuming that 1,000
file-descriptors are available, the offered load could be up to 500
requests per second (in practice, the sustainable load is often
somewhat smaller than the theoretical value).  On the downside,
timeouts artificially truncate the connection lifetime distribution.
Thus, it is recommended to pick a timeout value that is as large as
possible yet small enough to allow sustaining the desired offered
rate.  A timeout as short as one second may be acceptable, but larger
timeouts (5-10 seconds) are preferable.

It is important to keep in mind that timeouts do not guarantee that a
client can sustain a particular offered load---there are many other
potential resource bottlenecks.  For example, in some cases the client
machine may simply run out of CPU time.  To ensure that a given test
really measured the server's capabilities and not the client's, it is
a good idea to vary the number of machines participating in a test.
If observed performance remains the same as the number of client
machines is varied, the test results are likely to be valid.
.SH AUTHOR
.BR httperf
was developed by David Mosberger and was heavily influenced by an
earlier tool written by Tai Jin.  Stephane Eranian contributed the
log-file based URI generator.  Dick Carter contributed the
.B --wsesslog
workload generator, the support behind the
.B --period
option, and bug fixes.  All four authors are with Hewlett-Packard
Research Laboratories.
.SH BUGS
Probably many.  Always be sure to double-check results and don't fall
prey to measuring client-performance instead of server performance!
.PP
The user-interface definitely could be improved.  A simple workload
description language might be more suitable than the dozens of little
command-line options the tool has right now.