pmlogextract - Man Page

pmlogextract [-dfmwxz?] [-c configfile] [-S starttime] [-s samples] [-T endtime] [-V version] [-v volsize] [-Z timezone] input [...] output

pmlogextract reads one or more Performance Co-Pilot (PCP) archives identified by input and creates a merged and/or reduced PCP archive in output. Each input argument is either a name or a comma-separated list of names, and each name is the name of one file from an archive or the base name of an archive or the name of a directory containing one or more archives. The nature of merging is controlled by the number of input archives, while the nature of data reduction is controlled by the command line arguments. The input arguments must be archives created by pmlogger(1) with performance data collected from the same host, but usually over different time periods and possibly (although not usually) with different performance metrics being logged.

If only one input is specified, then the default behavior simply copies the input PCP archive (with possible conversion to a newer version of the archive format, see -V below), into the output PCP archive. When two or more PCP archives are specified as input, the archives are merged (or concatenated) and written to output.

In the output archive a <mark> record may be inserted at a time just past the end of each of the input archive to indicate a possible temporal discontinuity between the end of one input archive and the start of the next input archive. See the Mark Records section below for more information. There is no <mark> record after the end of the last (in temporal order) of the records from the input archive(s).

The available command line options are:

-c config, --config=config

Extract only the metrics specified in config from the input PCP archive(s). The config syntax accepted by pmlogextract is explained in more detail in the Configuration File Syntax section.

-d,  --desperate

Desperate mode. Normally if a fatal error occurs, all trace of the partially written PCP archive output is removed. With the -d option, the output archive is not removed.

-f,  --first

For most common uses, all of the input archives will have been collected in the same timezone. But if this is not the case, then pmlogextract must choose one of the timezones from the input archives to be used as the timezone for the output archive. The default is to use the timezone from the last input archive. The -f option forces the timezone from the first input archive to be used.

-m,  --mark

As described in the Mark Records section below, sometimes it is possible to safely omit <mark> records from the output archive. If the -m option is specified, then the epilogue and prologue test is skipped and a <mark> record will always be inserted at the end of each input archive (except the last). This is the original behaviour for pmlogextract.

-S starttime, --start=starttime

Define the start of a time window to restrict the records processed; refer to PCPIntro(1). See also the -w option.

-s samples, --samples=samples

The argument samples defines the number of samples (or records) to be written to output. If samples is 0 or -s is not specified, pmlogextract will continue until the end of all the input archives or until the end of the time window as specified by -T, whichever comes first. The -s option will override the -T option if it occurs sooner.

-T endtime, --finish=endtime

Define the end of a time window to restrict the records processed; refer to PCPIntro(1). See also the -w option.

-V version, --outputversion=version

Each PCP archive has a version for the physical record format, currently 2 or 3. By default, the output archive is created with a version equal to the maximum of the version of the input archives. The -V option may be used to explicitly force the version for output, provided version is no smaller than the archive version that would have been chosen by the default rule.

For example, specifying -V 3 may be used to produce a version 3 output archive from input archives that could be a mixture of version 2 and/or version 3.

-v volsize

The output archive is potentially a multi-volume data set, and the -v option causes pmlogextract to start a new volume after reaching an archive volume size of volsize. If volsize is an integer then at most volsize records will be written to each volume. If volsize is an integer suffixed by b or bytes then at most volsize bytes will be written out to each volume. Other viable file size units include: K, Kb, KiB, Kbyte, Kilobyte for kilobytes and M, Mb, MiB, Mbyte, Megabyte for megabytes and G, Gb, GiB, Gbyte, Gigabyte for gigabytes. These units may be optionally suffixed by an s and may be of mixed case. Alternatively volsize may be an integer or a floating point number suffixed using a time unit as described in PCPIntro(1) for the interval argument (to the standard PCP -t command line option).

Some examples of different formats:
-s 100
-s 100bytes
-s 100K
-s 100Mb
-s 10Gbyte
-s 10mins
-s 1.5hours

The default is for pmlogextract to produce as few volumes as possible.

Independent of any -v option, each volume of an archive is limited to no more than 2^31 bytes, so pmlogextract will automatically create a new volume for the archive before this limit is reached.

-w

Where -S and -T specify a time window within the same day, the -w flag will cause the data within the time window to be extracted, for every day in the archive. For example, the options -w -S @11:00 -T @15:00 specify that pmlogextract should include archive records only for the periods from 11am to 3pm on each day. When -w is used, the output archive will contain <mark> records to indicate the temporal discontinuity between the end of one time window and the start of the next.

-x

It is expected that the metadata (name, PMID, type, semantics and units) for each metric will be consistent across all of the input PCP archive(s) in which that metric appears. In rare cases, e.g. in development, in QA and when a PMDA is upgraded, this may not be the case and pmlogextract will report the issue and abort without creating the output archive. This is done so the problem can be fixed with pmlogrewrite(1) before retrying the merge. In unattended or QA environments it may be preferable to force the merge and omit the metrics with the mismatched metadata. The -x option does this.

-Z timezone, --timezone=timezone

Use timezone when displaying the date and time in diagnostics. Timezone is in the format of the environment variable TZ as described in environ(7). The default is to initially use the timezone of the local host.

-z,  --hostzone

Use the local timezone of the host from the input archive(s) when displaying the date and time in diagnostics. The default is to initially use the timezone of the local host.

-?,  --help

Display usage message and exit.

The configfile contains metrics of interest - only those metrics (or instances) mentioned explicitly or implicitly in the configuration file will be included in the output archive. Each specification must begin on a new line, and may span multiple lines in the configuration file. Instances may also be specified, but they are optional. The format for each specification is

        metric
or
        metric [ instance ... ]

where metric may be a leaf or a non-leaf name of a metric in the Performance Metrics Name Space (PMNS, see PMNS(5)). If a metric refers to a non-leaf node in the PMNS, pmlogextract will recursively descend the PMNS and include all metrics corresponding to descendent leaf nodes.

Instances are optional and are specified as a list space (or comma) separated of instance identifiers, with the list enclosed by square brackets. Each instance identifier may be a number or a string (enclosed in single or double quotes). instance identifiers that are numbers are assumed to be internal instance identifiers, else the string values are assumed to be external instance identifiers; see pmGetInDom(3) for more information. If no instances are given, then all instances of the associated metric(s) will be extracted.

Any additional white space is ignored and comments may be added with a `#' prefix.

This is an example of a valid configfile:

       #
       # config file for pmlogextract
       #

       kernel.all.cpu
       kernel.percpu.cpu.sys ["cpu0","cpu1"]
       disk.dev ["dks0d1"]

When more than one input archive contributes performance data to the output archive, then <mark> records may be inserted to indicate a possible temporal discontinuity in the performance data.

A <mark> record contains a timestamp and no performance data and is used to indicate that there is a time period in the PCP archive where we do not know the values of any performance metrics, because there was no pmlogger(1) collecting performance data during this period. Since these periods are often associated with the restart of a service or pmcd(1) or a system reboot, there may be considerable doubt as to the continuity of performance data across this time period.

Most current archives are created with a prologue record at the beginning and an epilogue record at the end. These records identify the state of pmcd(1) at the time, and may be used by pmlogextract to determine that there is no discontinuity between the end of one archive and the next output record, and as a consequence the <mark> record can safely be omitted from the output archive.

The rationale behind <mark> records may be demonstrated with an example. Consider one input archive that starts at 00:10 and ends at 09:15 on the same day, and another input archive that starts at 09:20 on the same day and ends at 00:10 the following morning. This would be a very common case for archives managed and rotated by pmlogger_check(1) and pmlogger_daily(1).

The output archive created by pmlogextract would contain:
00:10.000    first record from first input archive
...
09:15.000    last record from first input archive
09:15.001    <mark> record
09:20.000    first record from second input archive
...
01:10.000    last record from second input archive

The time period where the performance data is missing starts just after 09:15 and ends just before 09:20. When the output archive is processed with any of the PCP reporting tools, the <mark> record is used to indicate a period of missing data. For example using the output archive above, imagine one was reporting the average I/O rate at 30 minute intervals aligned on the hour and half-hour. The I/O count metric is a counter, so the average I/O rate requires two valid values from consecutive sample times. There would be values for all the intervals ending at 09:00, then no values at 09:30 because of the <mark> record, then no values at 10:00 because the “prior” value at 09:30 is not available, then the rate would be reported again at 10:30 and continue every 30 minutes until the last reported value at 01:00.

The presence of <mark> records in a PCP archive can be established using pmlogdump(1) where a timestamp and the annotation <mark> is used to indicate a <mark> record.

When more than one input archive is specified, pmlogextract performs a number of checks to ensure the metadata is consistent for metrics appearing in more than one of the input archives. These checks include:

• metric data type is the same
• metric semantics are the same
• metric units are the same
• metric is either always singular or always has the same instance domain
• metrics with the same name have the same PMID
• metrics with the same PMID have the same name

If any of these checks fail, pmlogextract reports the details and aborts without creating the output archive.

To address these semantic issues, use pmlogrewrite(1) to translate the input archives into equivalent archives with consistent metadata before using pmlogextract.

Refer to the -x and -d command line options above for alternatives to the default handling of errors during metadata checks.

The prologue metrics (pmcd.pmlogger.archive, pmcd.pmlogger.host, and pmcd.pmlogger.port), which are automatically recorded by pmlogger at the start of the archive, may not be present in the archive output by pmlogextract. These metrics are only relevant while the archive is being created, and have no significance once recording has finished.

All error conditions detected by pmlogextract are reported on stderr with textual (if sometimes terse) explanation.

If one of the input archives contains no archive records then an “empty archive” warning is issued and that archive is skipped.

Should one of the input archive(s) be corrupted (this can happen if the pmlogger instance writing the archive suddenly dies), then pmlogextract will detect and report the position of the corruption in the file, and any subsequent information from that archive will not be processed.

If any error is detected, pmlogextract will exit with a non-zero status.

For each of the input and output archive, several physical files are used.

archive.meta

metadata (metric descriptions, instance domains, etc.) for the archive

archive.0

initial volume of metrics values (subsequent volumes have suffixes 1, 2, ...) - for input these files may have been previously compressed with bzip2(1) or gzip(1) and thus may have an additional .bz2 or .gz suffix.

archive.index

temporal index to support rapid random access to the other files in the archive.

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).

For environment variables affecting PCP tools, see pmGetOptions(3).

PCPIntro(1), pmlc(1), pmlogdump(1), pmlogger(1), pmlogreduce(1), pmlogrewrite(1), pcp.conf(5), pcp.env(5) and PMNS(5).

ganglia2pcp(1), LOGARCHIVE(5), PCPIntro(1), pmFetch(3), pmFetchArchive(3), pmlogcheck(1), pmlogdump(1), pmlogger(1), pmlogger_daily(1), pmlogger_merge(1), pmloglabel(1), pmlogpaste(1), pmlogreduce(1), pmlogrewrite(1), pmlogsummary(1), pmrep(1), sar2pcp(1).