Chapter 2. Setting up PCP

Performance Co-Pilot (PCP) is a suite of tools, services, and libraries for monitoring, visualizing, storing, and analyzing system-level performance measurements.

This section describes how to install and enable PCP on your system.

2.1. Overview of PCP

You can add performance metrics using Python, Perl, C++, and C interfaces. Analysis tools can use the Python, C++, C client APIs directly, and rich web applications can explore all available performance data using a JSON interface.

You can analyze data patterns by comparing live results with archived data.

Features of PCP:

  • Light-weight distributed architecture, which is useful during the centralized analysis of complex systems.
  • It allows the monitoring and management of real-time data.
  • It allows logging and retrieval of historical data.

PCP has the following components:

  • The Performance Metric Collector Daemon (pmcd) collects performance data from the installed Performance Metric Domain Agents (pmda). PMDAs can be individually loaded or unloaded on the system and are controlled by the PMCD on the same host.
  • Various client tools, such as pminfo or pmstat, can retrieve, display, archive, and process this data on the same host or over the network.
  • The pcp package provides the command-line tools and underlying functionality.
  • The pcp-gui package provides the graphical application. Install the pcp-gui package by executing the yum install pcp-gui command. For more information, see Visually tracing PCP log archives with the PCP Charts application.

2.2. Installing and enabling PCP

To begin using PCP, install all the required packages and enable the PCP monitoring services.

This procedure describes how to install PCP using the pcp package. If you want to automate the PCP installation, install it using the pcp-zeroconf package. For more information on installing PCP by using pcp-zeroconf, see Setting up PCP with pcp-zeroconf.

Procedure

  1. Install the pcp package:

    # yum install pcp
  2. Enable and start the pmcd service on the host machine:

    # systemctl enable pmcd
    
    # systemctl start pmcd

Verification steps

  • Verify if the pmcd process is running on the host:

    # pcp
    
    Performance Co-Pilot configuration on workstation:
    
    platform: Linux workstation 4.18.0-80.el8.x86_64 #1 SMP Wed Mar 13 12:02:46 UTC 2019 x86_64
    hardware: 12 cpus, 2 disks, 1 node, 36023MB RAM
    timezone: CEST-2
    services: pmcd
    pmcd: Version 4.3.0-1, 8 agents
    pmda: root pmcd proc xfs linux mmv kvm jbd2

Additional resources

2.3. Deploying a minimal PCP setup

The minimal PCP setup collects performance statistics on Red Hat Enterprise Linux. The setup involves adding the minimum number of packages on a production system needed to gather data for further analysis.

You can analyze the resulting tar.gz file and the archive of the pmlogger output using various PCP tools and compare them with other sources of performance information.

Prerequisites

Procedure

  1. Update the pmlogger configuration:

    # pmlogconf -r /var/lib/pcp/config/pmlogger/config.default
  2. Start the pmcd and pmlogger services:

    # systemctl start pmcd.service
    
    # systemctl start pmlogger.service
  3. Execute the required operations to record the performance data.
  4. Stop the pmcd and pmlogger services:

    # systemctl stop pmcd.service
    
    # systemctl stop pmlogger.service
  5. Save the output and save it to a tar.gz file named based on the host name and the current date and time:

    # cd /var/log/pcp/pmlogger/
    
    # tar -czf $(hostname).$(date +%F-%Hh%M).pcp.tar.gz $(hostname)

    Extract this file and analyze the data using PCP tools.

Additional resources

2.4. System services distributed with PCP

The following table describes roles of various system services, which are distributed with PCP.

Table 2.1. Roles of system services distributed with PCP

Name

Description

pmcd

The Performance Metric Collector Daemon (PMCD).

pmie

The Performance Metrics Inference Engine.

pmlogger

The performance metrics logger.

pmproxy

The realtime and historical performance metrics proxy, time series query and REST API service.

2.5. Tools distributed with PCP

The following table describes usage of various tools, which are distributed with PCP.

Table 2.2. Usage of tools distributed with PCP

Name

Description

pcp

Displays the current status of a Performance Co-Pilot installation.

pcp-atop

Shows the system-level occupation of the most critical hardware resources from the performance point of view: CPU, memory, disk, and network.

pcp-atopsar

Generates a system-level activity report over a variety of system resource utilization. The report is generated from a raw logfile previously recorded using pmlogger or the -w option of pcp-atop.

pcp-dmcache

Displays information about configured Device Mapper Cache targets, such as: device IOPs, cache and metadata device utilization, as well as hit and miss rates and ratios for both reads and writes for each cache device.

pcp-dstat

Displays metrics of one system at a time. To display metrics of multiple systems, use --host option.

pcp-free

Reports on free and used memory in a system.

pcp-htop

Displays all processes running on a system along with their command line arguments in a manner similar to the top command, but allows you to scroll vertically and horizontally as well as interact using a mouse. You can also view processes in a tree format and select and act on multiple processes at once.

pcp-ipcs

Displays information on the inter-process communication (IPC) facilities that the calling process has read access for.

pcp-numastat

Displays NUMA allocation statistics from the kernel memory allocator.

pcp-pidstat

Displays information about individual tasks or processes running on the system such as: CPU percentage, memory and stack usage, scheduling, and priority. Reports live data for the local host by default.

pcp-ss

Displays socket statistics collected by the pmdasockets Performance Metrics Domain Agent (PMDA).

pcp-uptime

Displays how long the system has been running, how many users are currently logged on, and the system load averages for the past 1, 5, and 15 minutes.

pcp-vmstat

Provides a high-level system performance overview every 5 seconds. Displays information about processes, memory, paging, block IO, traps, and CPU activity.

pmchart

Plots performance metrics values available through the facilities of the Performance Co-Pilot.

pmclient

Displays high-level system performance metrics by using the Performance Metrics Application Programming Interface (PMAPI).

pmconfig

Displays the values of configuration parameters.

pmdbg

Displays available Performance Co-Pilot debug control flags and their values.

pmdiff

Compares the average values for every metric in either one or two archives, in a given time window, for changes that are likely to be of interest when searching for performance regressions.

pmdumplog

Displays control, metadata, index, and state information from a Performance Co-Pilot archive file.

pmdumptext

Outputs the values of performance metrics collected live or from a Performance Co-Pilot archive.

pmerr

Displays available Performance Co-Pilot error codes and their corresponding error messages.

pmfind

Finds PCP services on the network.

pmie

An inference engine that periodically evaluates a set of arithmetic, logical, and rule expressions. The metrics are collected either from a live system, or from a Performance Co-Pilot archive file.

pmieconf

Displays or sets configurable pmie variables.

pmiectl

Manages non-primary instances of pmie.

pminfo

Displays information about performance metrics. The metrics are collected either from a live system, or from a Performance Co-Pilot archive file.

pmiostat

Reports I/O statistics for SCSI devices (by default) or device-mapper devices (with the -x dm option).

pmlc

Interactively configures active pmlogger instances.

pmlogcheck

Identifies invalid data in a Performance Co-Pilot archive file.

pmlogconf

Creates and modifies a pmlogger configuration file.

pmlogctl

Manages non-primary instances of pmlogger.

pmloglabel

Verifies, modifies, or repairs the label of a Performance Co-Pilot archive file.

pmlogsummary

Calculates statistical information about performance metrics stored in a Performance Co-Pilot archive file.

pmprobe

Determines the availability of performance metrics.

pmrep

Reports on selected, easily customizable, performance metrics values.

pmsocks

Allows access to a Performance Co-Pilot hosts through a firewall.

pmstat

Periodically displays a brief summary of system performance.

pmstore

Modifies the values of performance metrics.

pmtrace

Provides a command line interface to the trace PMDA.

pmval

Displays the current value of a performance metric.

2.6. PCP deployment architectures

Performance Co-Pilot (PCP) offers many options to accomplish advanced setups. From the huge variety of possible architectures, this section describes how to scale your PCP deployment based on the recommended deployment set up by Red Hat, sizing factors, and configuration options.

PCP supports multiple deployment architectures, based on the scale of the PCP deployment.

Available scaling deployment setup variants:

Localhost

Each service runs locally on the monitored machine. When you start a service without any configuration changes, this is the default deployment. Scaling beyond the individual node is not possible in this case.

By default, the deployment setup for Redis is standalone, localhost. However, Redis can optionally perform in a highly-available and highly scalable clustered fashion, where data is shared across multiple hosts. Another viable option is to deploy a Redis cluster in the cloud, or to utilize a managed Redis cluster from a cloud vendor.

Decentralized

The only difference between localhost and decentralized setup is the centralized Redis service. In this model, the host executes pmlogger service on each monitored host and retrieves metrics from a local pmcd instance. A local pmproxy service then exports the performance metrics to a central Redis instance.

Figure 2.1. Decentralized logging

Decentralized logging
Centralized logging - pmlogger farm

When the resource usage on the monitored hosts is constrained, another deployment option is a pmlogger farm, which is also known as centralized logging. In this setup, a single logger host executes multiple pmlogger processes, and each is configured to retrieve performance metrics from a different remote pmcd host. The centralized logger host is also configured to execute the pmproxy service, which discovers the resulting PCP archives logs and loads the metric data into a Redis instance.

Figure 2.2. Centralized logging - pmlogger farm

Centralized logging - pmlogger farm
Federated - multiple pmlogger farms

For large scale deployments, Red Hat recommends to deploy multiple pmlogger farms in a federated fashion. For example, one pmlogger farm per rack or data center. Each pmlogger farm loads the metrics into a central Redis instance.

Figure 2.3. Federated - multiple pmlogger farms

Federated - multiple pmlogger farms
Note

By default, the deployment setup for Redis is standalone, localhost. However, Redis can optionally perform in a highly-available and highly scalable clustered fashion, where data is shared across multiple hosts. Another viable option is to deploy a Redis cluster in the cloud, or to utilize a managed Redis cluster from a cloud vendor.

Additional resources

2.8. Sizing factors

The following are the sizing factors required for scaling:

Remote system size
The number of CPUs, disks, network interfaces, and other hardware resources affects the amount of data collected by each pmlogger on the centralized logging host.
Logged Metrics
The number and types of logged metrics play an important role. In particular, the per-process proc.* metrics require a large amount of disk space, for example, with the standard pcp-zeroconf setup, 10s logging interval, 11 MB without proc metrics versus 155 MB with proc metrics - a factor of 10 times more. Additionally, the number of instances for each metric, for example the number of CPUs, block devices, and network interfaces also impacts the required storage capacity.
Logging Interval
The interval how often metrics are logged, affects the storage requirements. The expected daily PCP archive file sizes are written to the pmlogger.log file for each pmlogger instance. These values are uncompressed estimates. Since PCP archives compress very well, approximately 10:1, the actual long term disk space requirements can be determined for a particular site.
pmlogrewrite
After every PCP upgrade, the pmlogrewrite tool is executed and rewrites old archives if there were changes in the metric metadata from the previous version and the new version of PCP. This process duration scales linear with the number of archives stored.

Additional resources

  • pmlogrewrite(1) and pmlogger(1) man pages

2.9. Configuration options for PCP scaling

The following are the configuration options, which are required for scaling:

sysctl and rlimit settings
When archive discovery is enabled, pmproxy requires four descriptors for every pmlogger that it is monitoring or log-tailing, along with the additional file descriptors for the service logs and pmproxy client sockets, if any. Each pmlogger process uses about 20 file descriptors for the remote pmcd socket, archive files, service logs, and others. In total, this can exceed the default 1024 soft limit on a system running around 200 pmlogger processes. The pmproxy service in pcp-5.3.0 and later automatically increases the soft limit to the hard limit. On earlier versions of PCP, tuning is required if a high number of pmlogger processes are to be deployed, and this can be accomplished by increasing the soft or hard limits for pmlogger. For more information, see How to set limits (ulimit) for services run by systemd.
Local Archives
The pmlogger service stores metrics of local and remote pmcds in the /var/log/pcp/pmlogger/ directory. To control the logging interval of the local system, update the /etc/pcp/pmlogger/control.d/configfile file and add -t X in the arguments, where X is the logging interval in seconds. To configure which metrics should be logged, execute pmlogconf /var/lib/pcp/config/pmlogger/config.clienthostname. This command deploys a configuration file with a default set of metrics, which can optionally be further customized. To specify retention settings, that is when to purge old PCP archives, update the /etc/sysconfig/pmlogger_timers file and specify PMLOGGER_DAILY_PARAMS="-E -k X", where X is the amount of days to keep PCP archives.
Redis

The pmproxy service sends logged metrics from pmlogger to a Redis instance. The following are the available two options to specify the retention settings in the /etc/pcp/pmproxy/pmproxy.conf configuration file:

  • stream.expire specifies the duration when stale metrics should be removed, that is metrics which were not updated in a specified amount of time in seconds.
  • stream.maxlen specifies the maximum number of metric values for one metric per host. This setting should be the retention time divided by the logging interval, for example 20160 for 14 days of retention and 60s logging interval (60*60*24*14/60)

Additional resources

  • pmproxy(1), pmlogger(1), and sysctl(8) man pages

2.10. Example: Analyzing the centralized logging deployment

The following results were gathered on a centralized logging setup, also known as pmlogger farm deployment, with a default pcp-zeroconf 5.3.0 installation, where each remote host is an identical container instance running pmcd on a server with 64 CPU cores, 376 GB RAM, and one disk attached.

The logging interval is 10s, proc metrics of remote nodes are not included, and the memory values refer to the Resident Set Size (RSS) value.

Table 2.4. Detailed utilization statistics for 10s logging interval

Number of Hosts1050

PCP Archives Storage per Day

91 MB

522 MB

pmlogger Memory

160 MB

580 MB

pmlogger Network per Day (In)

2 MB

9 MB

pmproxy Memory

1.4 GB

6.3 GB

Redis Memory per Day

2.6 GB

12 GB

Table 2.5. Used resources depending on monitored hosts for 60s logging interval

Number of Hosts1050100

PCP Archives Storage per Day

20 MB

120 MB

271 MB

pmlogger Memory

104 MB

524 MB

1049 MB

pmlogger Network per Day (In)

0.38 MB

1.75 MB

3.48 MB

pmproxy Memory

2.67 GB

5.5GB

9 GB

Redis Memory per Day

0.54 GB

2.65 GB

5.3 GB

Note

The pmproxy queues Redis requests and employs Redis pipelining to speed up Redis queries. This can result in high memory usage. For troubleshooting this issue, see Troubleshooting high memory usage.

2.11. Example: Analyzing the federated setup deployment

The following results were observed on a federated setup, also known as multiple pmlogger farms, consisting of three centralized logging (pmlogger farm) setups, where each pmlogger farm was monitoring 100 remote hosts, that is 300 hosts in total.

This setup of the pmlogger farms is identical to the configuration mentioned in the Example: Analyzing the centralized logging deployment for 60s logging interval, except that the Redis servers were operating in cluster mode.

Table 2.6. Used resources depending on federated hosts for 60s logging interval

PCP Archives Storage per Daypmlogger MemoryNetwork per Day (In/Out)pmproxy MemoryRedis Memory per Day

277 MB

1058 MB

15.6 MB / 12.3 MB

6-8 GB

5.5 GB

Here, all values are per host. The network bandwidth is higher due to the inter-node communication of the Redis cluster.

2.12. Troubleshooting high memory usage

The following scenarios can result in high memory usage:

  • The pmproxy process is busy processing new PCP archives and does not have spare CPU cycles to process Redis requests and responses.
  • The Redis node or cluster is overloaded and cannot process incoming requests on time.

The pmproxy service daemon uses Redis streams and supports the configuration parameters, which are PCP tuning parameters and affects Redis memory usage and key retention. The /etc/pcp/pmproxy/pmproxy.conf file lists the available configuration options for pmproxy and the associated APIs.

This section describes how to troubleshoot high memory usage issue.

Prerequisites

  1. Install the pcp-pmda-redis package:

    # yum install pcp-pmda-redis
  2. Install the redis PMDA:

    # cd /var/lib/pcp/pmdas/redis && ./Install

Procedure

  • To troubleshoot high memory usage, execute the following command and observe the inflight column:

    $ pmrep :pmproxy
             backlog  inflight  reqs/s  resp/s   wait req err  resp err  changed  throttled
              byte     count   count/s  count/s  s/s  count/s   count/s  count/s   count/s
    14:59:08   0         0       N/A       N/A   N/A    N/A      N/A      N/A        N/A
    14:59:09   0         0    2268.9    2268.9    28     0        0       2.0        4.0
    14:59:10   0         0       0.0       0.0     0     0        0       0.0        0.0
    14:59:11   0         0       0.0       0.0     0     0        0       0.0        0.0

    This column shows how many Redis requests are in-flight, which means they are queued or sent, and no reply was received so far.

    A high number indicates one of the following conditions:

    • The pmproxy process is busy processing new PCP archives and does not have spare CPU cycles to process Redis requests and responses.
    • The Redis node or cluster is overloaded and cannot process incoming requests on time.
  • To troubleshoot the high memory usage issue, reduce the number of pmlogger processes for this farm, and add another pmlogger farm. Use the federated - multiple pmlogger farms setup.

    If the Redis node is using 100% CPU for an extended amount of time, move it to a host with better performance or use a clustered Redis setup instead.

  • To view the pmproxy.redis.* metrics, use the following command:

    $ pminfo -ftd pmproxy.redis
    pmproxy.redis.responses.wait [wait time for responses]
        Data Type: 64-bit unsigned int  InDom: PM_INDOM_NULL 0xffffffff
        Semantics: counter  Units: microsec
        value 546028367374
    pmproxy.redis.responses.error [number of error responses]
        Data Type: 64-bit unsigned int  InDom: PM_INDOM_NULL 0xffffffff
        Semantics: counter  Units: count
        value 1164
    [...]
    pmproxy.redis.requests.inflight.bytes [bytes allocated for inflight requests]
        Data Type: 64-bit int  InDom: PM_INDOM_NULL 0xffffffff
        Semantics: discrete  Units: byte
        value 0
    
    pmproxy.redis.requests.inflight.total [inflight requests]
        Data Type: 64-bit unsigned int  InDom: PM_INDOM_NULL 0xffffffff
        Semantics: discrete  Units: count
        value 0
    [...]

    To view how many Redis requests are inflight, see the pmproxy.redis.requests.inflight.total metric and pmproxy.redis.requests.inflight.bytes metric to view how many bytes are occupied by all current inflight Redis requests.

    In general, the redis request queue would be zero but can build up based on the usage of large pmlogger farms, which limits scalability and can cause high latency for pmproxy clients.

  • Use the pminfo command to view information about performance metrics. For example, to view the redis.* metrics, use the following command:

    $ pminfo -ftd redis
    redis.redis_build_id [Build ID]
        Data Type: string  InDom: 24.0 0x6000000
        Semantics: discrete  Units: count
        inst [0 or "localhost:6379"] value "87e335e57cffa755"
    redis.total_commands_processed [Total number of commands processed by the server]
        Data Type: 64-bit unsigned int  InDom: 24.0 0x6000000
        Semantics: counter  Units: count
        inst [0 or "localhost:6379"] value 595627069
    [...]
    
    redis.used_memory_peak [Peak memory consumed by Redis (in bytes)]
        Data Type: 32-bit unsigned int  InDom: 24.0 0x6000000
        Semantics: instant  Units: count
        inst [0 or "localhost:6379"] value 572234920
    [...]

    To view the peak memory usage, see the redis.used_memory_peak metric.

Additional resources