Performance Monitoring for the Cloud - Java2Days 2017

Copyright © 2016, Creative Arts & Technologies and others. All rights reserved.
Performance Monitoring
for the Cloud
Werner Keil
JSR 363 Maintenance Lead
@wernerkeil
October 18, 2017

© 2017 Creative Arts & Technologies and others. All rights reserved.#Monitoring #Performance
Agenda
1. Introduction
2. Performance Co-Pilot
3. Dropwizard Metrics
4. MicroProfile Metrics
5. Prometheus
6. StatsD
7. Demo

Who am I?
Werner Keil
• Consultant – Coach
• Creative Cosmopolitan
• Open Source Evangelist
• Software Architect
• Spec Lead – JSR363
• Individual JCP Executive Committee Member
[www.linkedin.com/in/catmedia]
Twitter @wernerkeil

What is Monitoring?
Monitoring applications is observing, analyzing and
manipulating the execution of these applications, which gives
information about threads, CPU usage, memory usage, as well
as other information like methods and classes being used.
A particular case is the monitoring of distributed
applications, aka the Cloud where an the performance
analysis of nodes and communication between them pose
additional challenges.

A high-level view of Cloud
Monitoring

Challenges at System Level
• Efficient Scalability
– Supporting tens of thousands of monitoring tasks
– Cost effective: minimize resource usage
• Monitoring QoS
– Multi-tenancy environment
– Minimize resource contention between monitoring tasks
• Implication of Multi-Tenancy
– Monitoring tasks: adding, removing
– Resource contention between monitoring tasks

Performance vs Number of Hosts
Number of hosts Performance (values per second)
100 100
1000 1000
10000 10000
60 items per host, update frequency once per minute
Number of hosts Performance (values per second)
100 1000
1000 10000
10000 100000
600 items per host, update frequency once per minute

Monitoring Tips
• Regularly apply “Little’s Law” to all data... generic
(queueing theory) form:
Q = λ R
• Length = Arrival Rate x Response Time
– e.g. 10 MB = 2 MB/sec x 5 sec
• Utilization = Arrival Rate x Service Time
– e.g. 20% = 0.2 = 100 msec/sec x 2 sec

Types of Monitoring
Monitoring Logs
• Logstash
• Redis
• Elasticsearch
• Kibana Dashboard
Monitoring Performance
• Collectd
• Statsd
• PCP
• Graphite
• Database (eg: PSQL)
• Grafana Dashboard

Monitoring Logs – Kibana
Dashboard

Monitoring Performance
How is this traditionally done?
• rsyslog/syslog-ng/journald
• top/iostat/vmstat/ps
• Mixture of scripting languages (bash/perl/python)
• Specific tools vary per platform
• Proper analysis requires more context

Performance Co-Pilot
PCP
http://www.pcp.io
GitHub
https://github.com/performancecopilot

What is PCP?
• Open source toolkit
• System-level analysis
• Live and historical
• Extensible (monitors, collectors)
• Distributed
• Unix-like component design
• Cross platform
• Ubiquitous units of measurement

PCP Basics
Agents and Daemons
At the core we have two basic
components:
1. Performance Metric
Domain Agents
• Agents
2. Performance Metric
Collection Daemon
• PMCD

PCP Architecture

PCP Metrics
• pminfo --desc -tT --fetch disk.dev.read
disk.dev.read [per-disk read operations]
Data Type: 32-bit unsigned int InDom: 60.1
Semantics: counter Units: count
Help: Cumulative count of disk reads since
boot time
Values:
inst [0 or "sda"] value 3382299
inst [1 or "sdb"] value 178421
16

PCP Agents
Webserver
(apache/nginx)
DBMS
Network
Kernel
PMCD

PCP Clients
Agents
PMCD
pmie
pmstat
pmval
pminfo
pmchart

PCP Remote Clients
Agents
PMCD
Clients
Remote
PMCD

PCP Data Model
• Metrics come from one source (host / archive)
• Source can be queried at any interval by any monitor tool
• Hierarchical metric names
e.g. disk.dev.read and aconex.response_time.avg
• Metrics are singular or set-valued (“instance domain”)
• Metadata associated with every metric
• Data type (int32, uint64, double, ...)
• Data semantics (units, scale, ...)
• Instance domain

Performance Timeline
• Where does the time go?
• Where’s it going now?
• Where will it go?

Performance Timeline – Toolkit
• Archives
• Live Monitoring
• Modelling and statistical
prediction

Performance Timeline – PCP
Toolkit
• Yesterday, last week, last month, ...
• All starts with pmlogger
• Arbitrary metrics, intervals
• One instance produces one PCP archive for one host
• An archive consists of 3 files
• Metadata, temporal index, data volume(s)
• pmlogger_daily, pmlogger_check
• Ensure the data keeps flowing
• pmlogsummary, pmwtf, pmdumptext
• pmlogextract, pmlogreduce

Custom Instrumentation
(Applications)

PCP – Parfait
Parfait has 4 main parts (for now)
• Monitoring
• DXM
• Timing
• Requests

Parfait – Monitoring
• This is the ‘original’ PCP bridge metrics (heavily modified)
• Simple Java objects (MonitoredValues) which wrap a value (e.g.
AtomicLong, String)
• MonitoredValues register themselves with a registry (container)
• When values changes, observers notice and output accordingly
• PCP
• JMX
• Other (Custom/Extended)
• Very simple to use
• ‘Default registry’ (legacy concept)

Parfait – Timing
• Logs the resources consumed by a request (an individual user action)
• Relies on a single request being thread-bound (and threads being used
exclusively)
• Basically needs a Map<Thread, Value>
• Take the value for a Thread at the start, and at the end
• Delta is the ‘cost’ of that request

Parfait – Timing Example
[2010-09-22 15:02:13,466 INFO ][ait.timing.Log4jSink][http-8080-Processor3
gedq93kl][192.168.7.132][20][] Top taskssummaryfeatures:tasks
taskssummaryfeatures:tasks Elapsed time: own 380.146316 ms, total
380.14688 ms Total CPU: own 150.0 ms, total 150.0 ms User CPU: own 140.0 ms,
total 140.0 ms System CPU: own 10.0 ms, total 10.0 ms Blocked count: own 40,
total 40 Blocked time: own 22 ms, total 22 ms Wait count: own 2, total 2
Wait time: own 8 ms, total 8 ms Database execution time: own 57 ms, total 57
ms Database execution count: own 11, total 11Database logical read count: own
0, total 0 Database physical read count: own 0, total 0 Database CPU time:
own 0 ms, total 0 ms Database received bytes: own 26188 By, total 26188 By
Database sent bytes: own 24868 By, total 24868 By Error Pages: own 0, total
0 Bobo execution time: own 40.742124 ms, total 40.742124 ms Bobo execution
count: own 2, total 2 Bytes transferred via bobo search: own 0 By, total 0 By
Super search entity count: own 0, total 0 Super search count: own 0, total 0
Bytes transferred via super search: own 0 By, total 0 By Elapsed time
during super search: own 0 ms, total 0 ms

Parfait – Requests
• As well as snapshotting requests after completion, for many metrics we
can see meaningful ‘in-progress’ values
• Simple JMX bean which ‘walks’ in-progress requests
• Tie in with ThreadContext (MDC abstraction)
• Include UserID
• ThreadID

PCP – Speed
Golang implementation of the PCP
instrumentation API
There are 3 main components
in the library
• Client
• Registry
• Metric

PCP – Speed Metric
• SingletonMetric
• This type defines a metric with no instance domain and only one
value. It requires type, semantics and unit for construction, and
optionally takes a couple of description strings.
A simple construction
metric, err := speed.NewPCPSingletonMetric(
42, // initial value
"simple.counter", // name
speed.Int32Type, // type
speed.CounterSemantics, // semantics
speed.OneUnit, // unit
"A Simple Metric", // short description
"This is a simple counter metric to demonstrate the speed API", // long desc
)

PCP for
Containers

PCP for Containers – Cgroup
Accounting
• [subsys].stat files below /sys/fs/cgroup
• individual cgroup or summed over children
• blkio
• IOPs/bytes, service/wait time – aggregate/per-dev
• Split up by read/write, sync/async
• cpuacct
• Processor use per-cgroup - aggregate/per-CPU
• memory
• mapped anon pages, page cache, writeback, swap, active/inactive LRU
state

PCP for Containers –
Namespaces
• Example: cat /proc/net/dev
• Contents differ inside vs outside a container
• Processes (e.g. cat) in containers run in different network, ipc, process,
uts, mount namespaces
• Namespaces are inherited across fork/clone
• Processes within a container share common view

PCP Container Analysis – Goals
• Allow targeting of individual containers
• e.g. /proc/net/dev
• pminfo --fetch network
• vs
• pminfo –fetch –container=crank network
• Zero installation inside containers required
• Simplify your life (dev_t auto-mapping)
• Data reduction (proc.*, cgroup.*)

PCP Container Analysis –
Mechanisms
• pminfo -f –host=acme.com –container=crank network
• Wire protocol extension
• Inform interested PCP collector agents
• Resolving container names, mapping names to cgroups, PIDs, etc.
• setns(2)
• Runs on the board, plenty of work remains
• New monitor tools with container awareness

What is Metrics?
• Code instrumentation
• Meters
• Gauges
• Counters
• Histograms
• Web app instrumentation
• Web app health check

Metrics Reporters
• Reporters
• Console
• CSV
• Slf4j
• JMX
• Advanced reporters
• Graphite
• Ganglia

Metrics 3rd Party Libraries
• AspectJ
• InfluxDB
• StatsD
• Cassandra
• Spring

Metrics Basics
• MetricsRegistry
• A collection of all the metrics for your application
• Usually one instance per JVM
• Use more in multi WAR deployment
• Names
• Each metric has a unique name
• Registry has helper methods for creating names
MetricRegistry.name(Queue.class, "items", "total")
//com.example.queue.items.total
MetricRegistry.name(Queue.class, "size", "byte")
//com.example.queue.size.byte

Metrics Elements
• Gauges
• The simplest metric type: it just returns a value
final Map<String, String> keys = new HashMap<>();
registry.register(MetricRegistry.name("gauge", "keys"), new
Gauge<Integer>() {
@Override
public Integer getValue() {
return keys.keySet().size();
}
});

Metrics Elements (2)
• Counters
• Incrementing and decrementing 64.bit integer
final Counter counter= registry.counter(MetricRegistry.name("counter",
"inserted"));
counter.inc();

• Histograms
• Measures the distribution of values in a stream of data
final Histogram resultCounts = registry.histogram(name(ProductDAO.class,
"result-counts");
resultCounts.update(results.size());
• Meters
• Measures the rate at which a set of events occur
final Meter meter = registry.meter(MetricRegistry.name("meter", "inserted"));
meter.mark();

• Timers
• A histogram of the duration of a type of event and a meter of the rate
of its occurrence
Timer timer = registry.timer(MetricRegistry.name("timer", "inserted"));
Context context = timer.time();
//timed ops
context.stop();

Metrics – Graphite Reporter
final Graphite graphite = new Graphite(new
InetSocketAddress("graphite.example.com", 2003));
final GraphiteReporter reporter = GraphiteReporter.forRegistry(registry)
.prefixedWith("web1.example.com")
.convertRatesTo(TimeUnit.SECONDS)
.convertDurationsTo(TimeUnit.MILLISECONDS)
.filter(MetricFilter.ALL)
.build(graphite);
reporter.start(1, TimeUnit.MINUTES);
Metrics can be prefixed
Useful to divide environment metrics: prod, test

Metrics – Grafana Application
Overview

What is Eclipse MicroProfile?
● Eclipse MicroProfile is an open-source community
specification for Enterprise Java microservices
● A community of individuals, organizations, and vendors
collaborating within an open source (Eclipse) project to
bring microservices to the Enterprise Java community
47

Specifications 1.2
48
MicroProfile 1.2
= New
= No change from last release
JAX-RS 2.0JSON-P 1.0CDI 1.2
Config 1.1
Fault
Tolerance
1.0
JWT
Propagation
1.0
Health
Check 1.0
Metrics 1.0

Prometheus

What is Prometheus?
Prometheus is an open-source systems monitoring
and alerting toolkit originally
built at SoundCloud. It is now a standalone open
source project and maintained
independently of any company.

Prometheus Components
• The main Prometheus server which scrapes and stores time series data
• Client libraries for instrumenting application code
• A push gateway for supporting short-lived jobs
• Special-purpose exporters (for HAProxy, StatsD, Graphite, etc.)
• An alertmanager
• Various support tools
• WhiteBox Monitoring instead of probing (aka BlackBox Monitoring)

What is StatsD?
A network daemon that runs on the Node.js platform
and listens for statistics, like counters and timers, sent
over UDP or TCP and sends aggregates to one or more
pluggable backend services (e.g., Graphite).
StatsD was inspired (heavily) by the project (of the
same name) at Flickr.

Links
Performance Co-Pilot
http://www.pcp.io
Dropwizard Metrics
http://metrics.dropwizard.io
Eclipse MicroProfile
http://microprofile.io
Prometheus
http://prometheus.io
StatsD
https://github.com/etsy/statsd/wiki

Performance Monitoring for the Cloud - Java2Days 2017

More Related Content

What's hot (20)

Similar to Performance Monitoring for the Cloud - Java2Days 2017 (20)

More from Werner Keil (20)

Recently uploaded (20)

Performance Monitoring for the Cloud - Java2Days 2017

Editor's Notes