NAME¶

pmie - inference engine for performance metrics

SYNOPSIS¶

pmie [-bCdefHqVvWxz] [-A align] [-a archive] [-c filename] [-h host] [-l logfile] [-j stompfile] [-n pmnsfile] [-O offset] [-S starttime] [-T endtime] [-t interval] [-U username] [-Z timezone] [filename ...]

DESCRIPTION¶

pmie accepts a collection of arithmetic, logical, and rule expressions to be evaluated at specified frequencies. The base data for the expressions consists of performance metrics values delivered in real-time from any host running the Performance Metrics Collection Daemon (PMCD), or using historical data from Performance Co-Pilot (PCP) archive logs.

As well as computing arithmetic and logical values, pmie can execute actions (popup alarms, write system log messages, and launch programs) in response to specified conditions. Such actions are extremely useful in detecting, monitoring and correcting performance related problems.

The expressions to be evaluated are read from configuration files specified by one or more filename arguments. In the absence of any filename, expressions are read from standard input.

A description of the command line options specific to pmie follows:

-a

archive which is a comma-separated list of names, each of which may be the base name of an archive or the name of a directory containing one or more archives written by pmlogger(1). Multiple instances of the -a flag may appear on the command line to specify a list of sets of archives. In this case, it is required that only one set of archives be present for any one host. Also, any explicit host names occurring in a pmie expression must match the host name recorded in one of the archive labels. In the case of multiple sets of archives, timestamps recorded in the archives are used to ensure temporal consistency.

-b

Output will be line buffered and standard output is attached to standard error. This is most useful for background execution in conjunction with the -l option. The -b option is always used for pmie instances launched from pmie_check(1).

-C

Parse the configuration file(s) and exit before performing any evaluations. Any errors in the configuration file are reported.

-c

An alternative to specifying filename at the end of the command line.

-d

Normally pmie would be launched as a non-interactive process to monitor and manage the performance of one or more hosts. Given the -d flag however, execution is interactive and the user is presented with a menu of options. Interactive mode is useful mainly for debugging new expressions.

-e

When used with -V, -v or -W, this option forces timestamps to be reported with each expression. The timestamps are in ctime(3) format, enclosed in parenthesis and appear after the expression name and before the expression value, e.g.

	expr_1 (Tue Feb  6 19:55:10 2001): 12

-f

If the -l option is specified and there is no -a option (ie. real-time monitoring) then pmie is run as a daemon in the background (in all other cases foreground is the default). The -f option forces pmie to be run in the foreground, independent of any other options.

-h

By default performance data is fetched from the local host (in real-time mode) or the host for the first named set of archives on the command line (in archive mode). The host argument overrides this default. It does not override hosts explicitly named in the expressions being evaluated. The host argument is interpreted as a connection specification for pmNewContext, and is later mapped to the remote pmcd's self-reported host name for reporting purposes. See also the %h vs. %c substitutions in rule action strings below.

-l

Standard error is sent to logfile.

-j

An alternative STOMP protocol configuration is loaded from stompfile. If this option is not used, and the stomp action is used in any rule, the default location $PCP_SYSCONF_DIR/pmie/config/stomp will be used.

-n

An alternative Performance Metrics Name Space (PMNS) is loaded from the file pmnsfile.

-P

Identifies this as the primary pmie instance for a host. See the ``AUTOMATIC RESTART'' section below for further details.

-q

Suppresses diagnostic messages that would be printed to standard output by default, especially the "evaluator exiting" message as this can confuse scripts.

-t

The interval argument follows the syntax described in PCPIntro(1), and in the simplest form may be an unsigned integer (the implied units in this case are seconds). The value is used to determine the sample interval for expressions that do not explicitly set their sample interval using the pmie variable delta described below. The default is 10.0 seconds.

-U username

User account under which to run pmie. The default is the current user account for interactive use. When run as a daemon, the unprivileged "pcp" account is used in current versions of PCP, but in older versions the superuser account ("root") was used by default.

-v

Unless one of the verbose options -V, -v or -W appears on the command line, expressions are evaluated silently, the only output is as a result of any actions being executed. In the verbose mode, specified using the -v flag, the value of each expression is printed as it is evaluated. The values are in canonical units; bytes in the dimension of ``space'', seconds in the dimension of ``time'' and events in the dimension of ``count''. See pmLookupDesc(3) for details of the supported dimension and scaling mechanisms for performance metrics. The verbose mode is useful in monitoring the value of given expressions, evaluating derived performance metrics, passing these values on to other tools for further processing and in debugging new expressions.

-V

This option has the same effect as the -v option, except that the name of the host and instance (if applicable) are printed as well as expression values.

-W

This option has the same effect as the -V option described above, except that for boolean expressions, only those names and values that make the expression true are printed. These are the same names and values accessible to rule actions as the %h, %i, %c and %v bindings, as described below.

-x

Execute in domain agent mode. This mode is used within the Performance Co-Pilot product to derive values for summary metrics, see pmdasummary(1). Only restricted functionality is available in this mode (expressions with actions may not be used).

-Z

Change the reporting timezone to timezone in the format of the environment variable TZ as described in environ(7).

-z

Change the reporting timezone to the timezone of the host that is the source of the performance metrics, as identified via either the -h option or the first named set of archives (as described above for the -a option).

The -S, -T, -O, and -A options may be used to define a time window to restrict the samples retrieved, set an initial origin within the time window, or specify a ``natural'' alignment of the sample times; refer to PCPIntro(1) for a complete description of these options.

Output from pmie is directed to standard output and standard error as follows:

stdout

Expression values printed in the verbose -v mode and the output of print actions.

stderr

Error and warning messages for any syntactic or semantic problems during expression parsing, and any semantic or performance metrics availability problems during expression evaluation.

EXAMPLES¶

The following example expressions demonstrate some of the capabilities of the inference engine.

The directory $PCP_DEMOS_DIR/pmie contains a number of other annotated examples of pmie expressions.

The variable delta controls expression evaluation frequency. Specify that subsequent expressions be evaluated once a second, until further notice:

delta = 1 sec;

If the total context switch rate exceeds 10000 per second per CPU, then display an alarm notifier:

kernel.all.pswitch / hinv.ncpu > 10000 count/sec
-> alarm "high context switch rate %v";

If the high context switch rate is sustained for 10 consecutive samples, then launch top(1) in an xterm(1) window to monitor processes, but do this at most once every 5 minutes:

all_sample (


    kernel.all.pswitch @0..9 > 10 Kcount/sec * hinv.ncpu
) -> shell 5 min "xterm -e 'top'";

The following rules are evaluated once every 20 seconds:

delta = 20 sec;

If any disk is performing more than 60 I/Os per second, then print a message identifying the busy disk to standard output and launch dkvis(1):

some_inst (


    disk.dev.total > 60 count/sec
) -> print "busy disks:" " %i" &


     shell 5 min "dkvis";

Refine the preceding rule to apply only between the hours of 9am and 5pm, and to require 3 of 4 consecutive samples to exceed the threshold before executing the action:

$hour >= 9 && $hour <= 17 &&
some_inst (


  75 %_sample (


    disk.dev.total @0..3 > 60 count/sec


  )
) -> print "disks busy for 20 sec:" " [%h]%i";

The following two rules are evaluated once every 10 minutes:

delta = 10 min;

If either the / or the /usr filesystem is more than 95% full, display an alarm popup, but not if it has already been displayed during the last 4 hours:

filesys.free #'/dev/root' /


    filesys.capacity #'/dev/root' < 0.05
-> alarm 4 hour "root filesystem (almost) full";
filesys.free #'/dev/usr' /


    filesys.capacity #'/dev/usr' < 0.05
-> alarm 4 hour "/usr filesystem (almost) full";

The following rule requires a machine that supports the lmsensors metrics. If the machine environment temperature rises more than 2 degrees over a 10 minute interval, write an entry in the system log:

lmsensors.coretemp_isa.temp1 @0 - lmsensors.coretemp_isa.temp1 @1 > 2
-> alarm "temperature rising fast" &


   syslog "machine room temperature rise alarm";

And something interesting if you have performance problems with your Oracle database:

// back to 30sec evaluations
delta = 30 sec;
sid = "ptg1";		# $ORACLE_SID setting
lid = "223";		# latch ID from v$latch
lru = "#'$sid/$lid cache buffers lru chain'";
host = ":moomba.melbourne.sgi.com";
gets = "oracle.latch.gets $host $lru";
total = "oracle.latch.gets $host $lru +


         oracle.latch.misses $host $lru +


         oracle.latch.immisses $host $lru";
$total > 100 && $gets / $total < 0.2
-> alarm "high lru latch contention in database $sid";

The following ruleset will emit exactly one message depending on the availability and value of the 1-minute load average.

delta = 1 minute;
ruleset


     kernel.all.load #'1 minute' > 10 * hinv.ncpu ->


         print "extreme load average %v"
else kernel.all.load #'1 minute' > 2 * hinv.ncpu ->


         print "moderate load average %v"
unknown ->


         print "load average unavailable"
otherwise ->


         print "load average OK"
;

The following rule will emit a message when some filesystem is more than 75% full and is filling at a rate that if sustained would fill the filesystem to 100% in less than 30 minutes.

some_inst (


    100 * filesys.used / filesys.capacity > 75 &&


    filesys.used + 30min * (rate filesys.used) > filesys.capacity
) -> print "filesystem will be full within 30 mins:" " %i";

If the metric mypmda.errors counts errors then the following rule will emit a message if the rate of errors exceeds 1 per second provided the error count is less than 100.

mypmda.errors > 1 && instant mypmda.errors < 100
-> print "high error rate: %v";

QUICK START¶

The pmie specification language is powerful and large.

To expedite rapid development of pmie rules, the pmieconf(1) tool provides a facility for generating a pmie configuration file from a set of generalized pmie rules. The supplied set of rules covers a wide range of performance scenarios.

The Performance Co-Pilot User's and Administrator's Guide provides a detailed tutorial-style chapter covering pmie.

EXPRESSION SYNTAX¶

This description is terse and informal. For a more comprehensive description see the Performance Co-Pilot User's and Administrator's Guide.

A pmie specification is a sequence of semicolon terminated expressions.

Basic operators are modeled on the arithmetic, relational and Boolean operators of the C programming language. Precedence rules are as expected, although the use of parentheses is encouraged to enhance readability and remove ambiguity.

Operands are performance metric names (see pmns(5)) and the normal literal constants.

Operands involving performance metrics may produce sets of values, as a result of enumeration in the dimensions of hosts, instances and time. Special qualifiers may appear after a performance metric name to define the enumeration in each dimension. For example,

kernel.percpu.cpu.user :foo :bar #cpu0 @0..2

defines 6 values corresponding to the time spent executing in user mode on CPU 0 on the hosts ``foo'' and ``bar'' over the last 3 consecutive samples. The default interpretation in the absence of : (host), # (instance) and @ (time) qualifiers is all instances at the most recent sample time for the default source of PCP performance metrics.

Host and instance names that do not follow the rules for variables in programming languages, ie. alphabetic optionally followed by alphanumerics, should be enclosed in single quotes.

Expression evaluation follows the law of ``least surprises''. Where performance metrics have the semantics of a counter, pmie will automatically convert to a rate based upon consecutive samples and the time interval between these samples. All numeric expressions are evaluated in double precision, and where appropriate, automatically scaled into canonical units of ``bytes'', ``seconds'' and ``counts''.

A rule is a special form of expression that specifies a condition or logical expression, a special operator (->) and actions to be performed when the condition is found to be true.

The following table summarizes the basic pmie operators:

All operators are supported for numeric-valued operands and expressions. For string-valued operands, namely literal string constants enclosed in double quotes or metrics with a data type of string (PM_TYPE_STRING), only the operators == and != are supported.

The rate and instant operators are the logical inverse of one another, so an arithmetic expression expr is equal to rate instant expr. The more useful cases involve using rate with a metric that is not a counter to determine the rate of change over time or instant with a metric that is a counter to determine if the current value is above or below some threshold.

Aggregate operators may be used to aggregate or summarize along one dimension of a set-valued expression. The following aggregate operators map from a logical expression to a logical expression of lower dimension.

The following instantial operators may be used to filter or limit a set-valued logical expression, based on regular expression matching of instance names. The logical expression must be a set involving the dimension of instances, and the regular expression is of the form used by egrep(1) or the Extended Regular Expressions of regcomp(3).

For example, the expression below will be ``true'' for disks attached to controllers 2 or 3 performing more than 20 operations per second:

match_inst "^dks[23]d" disk.dev.total > 20;

The following aggregate operators map from an arithmetic expression to an arithmetic expression of lower dimension.

The aggregate operators count_inst, count_host and count_sample map from a logical expression to an arithmetic expression of lower dimension by counting the number of set members for which the expression is true in the associated dimension.

For action rules, the following actions are defined:

Multiple actions may be separated by the & and | operators to specify respectively sequential execution (both actions are executed) and alternate execution (the second action will only be executed if the execution of the first action returns a non-zero error status.

Arguments to actions are an optional suppression time, and then one or more expressions (a string is an expression in this context). Strings appearing as arguments to an action may include the following special selectors that will be replaced at the time the action is executed.

%h

Host name(s) that make the left-most top-level expression in the condition true.

%c

Connection specification string(s) or files for a PCP tool to reach the hosts or archives that make the left-most top-level expression in the condition true.

%i

Instance(s) that make the left-most top-level expression in the condition true.

%v

One value from the left-most top-level expression in the condition for each host and instance pair that makes the condition true.

Note that expansion of the special selectors is done by repeating the whole argument once for each unique binding to any of the qualifying special selectors. For example if a rule were true for the host mumble with instances grunt and snort, and for host fumble the instance puff makes the rule true, then the action

...
-> shell myscript "Warning: %h:%i busy ";

By comparison, if the action

...
-> shell myscript "Warning! busy:" " %h:%i";

The semantics of the expansion of the special selectors leads to a common usage pattern in an action, where one argument is a constant (contains no special selectors) the second argument contains the desired special selectors with minimal separator characters, and an optional third argument provides a constant postscript (e.g. to terminate any argument quoting from the first argument). If necessary post-processing (eg. in myscript) can provide the necessary enumeration over each unique expansion of the string containing just the special selectors.

For complex conditions, the bindings to these selectors is not obvious. It is strongly recommended that pmie be used in the debugging mode (specify the -W command line option in particular) during rule development.

BOOLEAN EXPRESSIONS¶

pmie expressions that have the semantics of a Boolean, e.g. foo.bar > 10 or some_inst ( my.table < 0 ) are assigned the values true or false or unknown. A value is unknown if one or more of the underlying metric values is unavailable, e.g. pmcd(1) on the host cannot be contacted, the metric is not in the PCP archive, no values are currently available, insufficient values have been fetched to allow a rate converted value to be computed or insufficient values have been fetched to instantiate the required number of samples in the temporal domain.

Boolean operators follow the normal rules of Kleene logic (aka 3-valued logic) when combining values that include unknown:

RULESETS¶

The ruleset clause is used to define a set of rules and actions that are evaluated in order until some action is executed, at which point the remaining rules and actions are skipped until the ruleset is again scheduled for evaluation. The keyword else is used to separate rules. After one or more regular rules (with a predicate and an action), a ruleset may include an optional
unknown -> action
clause, optionally followed by a
otherwise -> action
clause.

If all of the predicates in the rules evaluate to unknown and an unknown clause has been specified then action associated with the unknown clause will be executed.

If no rule predicate is true and the unknown action is either not specified or not executed and an otherwise clause has been specified, then the action associated with the otherwise clause will be executed.

SCALE FACTORS¶

Scale factors may be appended to arithmetic expressions and force linear scaling of the value to canonical units. Simple scale factors are constructed from the keywords: nanosecond, nanosec, nsec, microsecond, microsec, usec, millisecond, millisec, msec, second, sec, minute, min, hour, byte, Kbyte, Mbyte, Gbyte, Tbyte, count, Kcount and Mcount, and the operator /, for example ``Kbytes / hour''.

MACROS¶

Macros are defined using expressions of the form:

name = constexpr;

Where name follows the normal rules for variables in programming languages, ie. alphabetic optionally followed by alphanumerics. constexpr must be a constant expression, either a string (enclosed in double quotes) or an arithmetic expression optionally followed by a scale factor.

Macros are expanded when their name, prefixed by a dollar ($) appears in an expression, and macros may be nested within a constexpr string.

The following reserved macro names are understood.

minute

Current minute of the hour.

hour

Current hour of the day, in the range 0 to 23.

day

Current day of the month, in the range 1 to 31.

month

Current month of the year, in the range 0 (January) to 11 (December).

year

Current year.

day_of_week

Current day of the week, in the range 0 (Sunday) to 6 (Saturday).

delta

Sample interval in effect for this expression.

Dates and times are presented in the reporting time zone (see description of -Z and -z command line options above).

AUTOMATIC RESTART¶

It is often useful for pmie processes to be started and stopped when the local host is booted or shutdown, or when they have been detected as no longer running (when they have unexpectedly exited for some reason). Refer to pmie_check(1) for details on automating this process.

Optionally, each system running pmcd(1) may also be configured to run a ``primary'' pmie instance. This pmie instance is launched by $PCP_RC_DIR/pmie, and is affected by the files $PCP_SYSCONF_DIR/pmie/control, $PCP_SYSCONF_DIR/pmie/control.d (use chkconfig(8), systemctl(1) or similar platform-specific commands to activate or disable the primary pmie instance) and $PCP_VAR_DIR/config/pmie/config.default (the default initial configuration file for the primary pmie).

The primary pmie instance is identified by the -P option. There may be at most one ``primary'' pmie instance on each system. The primary pmie instance (if any) must be running on the same host as the pmcd(1) to which it connects (if any), so the -h and -P options are mutually exclusive.

EVENT MONITORING¶

It is common for production systems to be monitored in a central location. Traditionally on UNIX systems this has been performed by the system log facilities - see logger(1), and syslogd(1). On Windows, communication with the system event log is handled by pcp-eventlog(1).

pmie fits into this model when rules use the syslog action. Note that if the action string begins with -p (priority) and/or -t (tag) then these are extracted from the string and treated in the same way as in logger(1) and pcp-eventlog(1).

However, it is common to have other event monitoring frameworks also, into which you may wish to incorporate performance events from pmie. You can often use the shell action to send events to these frameworks, as they usually provide their a program for injecting events into the framework from external sources.

A final option is use of the stomp (Streaming Text Oriented Messaging Protocol) action, which allows pmie to connect to a central JMS (Java Messaging System) server and send events to the PMIE topic. Tools can be written to extract these text messages and present them to operations people (via desktop popup windows, etc). Use of the stomp action requires a stomp configuration file to be setup, which specifies the location of the JMS server host, port number, and username/password.

The format of this file is as follows:

host=messages.sgi.com   # this is the JMS server (required)
port=61616              # and its listening here (required)
timeout=2               # seconds to wait for server (optional)
username=joe            # (required)
password=j03ST0MP       # (required)
topic=PMIE              # JMS topic for pmie messages (optional)

The timeout value specifies the time (in seconds) that pmie should wait for acknowledgements from the JMS server after sending a message (as required by the STOMP protocol). Note that on startup, pmie will wait indefinitely for a connection, and will not begin rule evaluation until that initial connection has been established. Should the connection to the JMS server be lost at any time while pmie is running, pmie will attempt to reconnect on each subsequent truthful evaluation of a rule with a stomp action, but not more than once per minute. This is to avoid contributing to network congestion. In this situation, where the STOMP connection to the JMS server has been severed, the stomp action will return a non-zero error value.

FILES¶

$PCP_DEMOS_DIR/pmie/*

annotated example rules

$PCP_VAR_DIR/pmns/*

default PMNS specification files

$PCP_TMP_DIR/pmie

pmie maintains files in this directory to identify the running pmie instances and to export runtime information about each instance - this data forms the basis of the pmcd.pmie performance metrics

$PCP_PMIECONTROL_PATH

the default set of pmie instances to start at boot time - refer to pmie_check(1) for details

BUGS¶

The lexical scanner and parser will attempt to recover after an error in the input expressions. Parsing resumes after skipping input up to the next semi-colon (;), however during this skipping process the scanner is ignorant of comments and strings, so an embedded semi-colon may cause parsing to resume at an unexpected place. This behavior is largely benign, as until the initial syntax error is corrected, pmie will not attempt any expression evaluation.

PCP ENVIRONMENT¶

Environment variables with the prefix PCP_ are used to parameterize the file and directory names used by PCP. On each installation, the file /etc/pcp.conf contains the local values for these variables. The $PCP_CONF variable may be used to specify an alternative configuration file, as described in pcp.conf(5).

When executing shell actions, pmie overrides two variables - IFS and PATH - in the environment of the child process. IFS is set to "\t\n". The PATH is set to a combination of a default path for all platforms ("/usr/sbin:/sbin:/usr/bin:/usr/sbin") and several configurable components. These are (in this order): $PCP_BIN_DIR, $PCP_BINADM_DIR and $PCP_PLATFORM_PATHS.

When executing popup alarm actions, pmie will use the value of $PCP_XCONFIRM_PROG as the visual notification program to run. This is typically set to pmconfirm(1), a cross-platform dialog box.

UNIX SEE ALSO¶

logger(1).

WINDOWS SEE ALSO¶

pcp-eventlog(1).

SEE ALSO¶

PCPIntro(1), pmcd(1), pmconfirm(1), pmdumplog(1), pmieconf(1), pmie_check(1), pminfo(1), pmlogger(1), pmval(1), PMAPI(3), pcp.conf(5) and pcp.env(5).

USER GUIDE¶

For a more complete description of the pmie language, refer to the Performance Co-Pilot Users and Administrators Guide. This is available online from:

https://pcp.io/doc/pcp-users-and-administrators-guide.pdf