12.2 Oracle QoS Management TWP
Oracle Database 12c Release 2:
Quality of Service Management
Monitoring and Managing Oracle RAC Database Performance
ORAC LE WHI TE PAPER
|
MARCH 2017
Table of Contents
Introduction
2
Datacenter Runtime Management Requirements
3
Runtime Management Best Practices – The Phases
4
Phase 1: Plan the Deployment
4
Phase 2: Runtime Measure the Deployment
7
Phase 3: Runtime Monitor the Deployment
10
Phase 4: Runtime Manage the Deployment
12
Conclusion
15
Appendix
16
1
Introduction
The database is no longer the center of the universe. Such a statement would have been heretical
just a short time ago. However, the introduction of the on-premise database cloud and DBaaS has
altered the deployment strategy and database management requirements. It is no longer sufficient
to plan for simply performance, scalability and high availability. These new deployment models must
also consider consolidation, provisioning, patching and quality of service. Oracle Real Application
Cluster databases whether in single node form as RAC One Node, Standard multi-node form, or the
new 12c multitenant form, provide the level of performance, availability and manageability to be the
foundation of modern consolidated on-premise database clouds or Database-as-a-Service
deployments.
Figure 1: Evolution of Oracle RAC Database
The ability to manage complex highly available database service deployments in real time is now a
common requirement as enterprises adopt a database service-centric deployment model where
multiple databases share common physical resources and are no longer silo’d on dedicated
hardware. Where resource utilization has improved and IT spend optimized, runtime management
complexity has increased. Oracle has addressed this in the Oracle 12c RAC release with new
Oracle Database Quality of Service Management (QoS) functionality to support all deployment
types.
2
Figure 2: On-Premise Database Cloud Runtime Management
This functionality is included in the Oracle RAC and RAC One Node license and its management
interface is integrated into the Enterprise Manager Cloud Control Database Plug-in. Therefore, no
additional management packs are required.
Datacenter Runtime Management Requirements
The basic tenants under which Oracle Database QoS Management was developed can be distilled into three
statements that must be able to be achieved in real time:
» When resources across the datacenter are sufficient they are continuously deployed to ensure performance
and availability objectives will be met.
» When resources are insufficient to meet demand more business-critical objectives will be met at the expense
of less critical ones.
» When load conditions severely exceed capacity, resources remain available.
In order to achieve these goals, specific functionality must be built into the entire software stack to include
accurate measurement of performance, resource bottleneck analysis, resource trade-off evaluation and online
dynamic resource allocation.
In the end, the effectiveness of achieving the above goals is evaluated by each application’s performance over
time. When examining modern multi-tier applications, it should not be unexpected that most of a transaction’s
response time is contained in the database tier and its associated storage. This performance can be distilled at a
high level into the following simple equation:
Resource Use + Resource Wait = Application Performance
It’s important to realize that once an application is deployed, there is almost no ability for online management of
its use of resources, as these were the responsibility of design, development, Q/A, and test teams. However,
3
there is the potential for the online management of the amount of time needed to wait for resources, whether
these are CPU, memory, or I/O.
Fortunately, the Oracle software stack, especially the database tier, has rich resource management capabilities
that have been enhanced in Oracle 12c to facilitate this when used in concert with QoS Management.
Runtime Management Best Practices – The Phases
The best practices for runtime management of an Oracle RAC-based on-premise database cloud or Databaseas-a-Service deployment may be applied in discrete phases to gain insight into the actual workloads and their
use of resources as well as confidence in setting realistic service level agreements (SLAs) and the ability to
manage to them. The four phases that will be discussed are as follows:
1.
Plan the deployment
2.
Runtime measure the deployment
3.
Runtime monitor the deployment
4.
Runtime manage the deployment
These phases should be implemented serially, and not combined to accelerate deployment as each captures
necessary data that is used in the next phase.
Phase 1: Plan the Deployment
Planning the deployment may start at various points, but for this paper we will assume the deployment is to be a
on-premise database cloud offering database services to applications each of which has an importance or
criticality to the business that may vary due to calendar or events. This paper is not intended to focus on this
particular task, but will introduce its elements.
Since the introduction of Oracle Database 11.2, customers have had a choice of three different cluster database
deployment types – administrator-managed, policy-managed or a hybrid of the two. While it is beyond the scope
of this paper to explore the pros and cons of each type, as a general rule if the databases to be deployed are
11.2 or greater, then policy-managed should be fully evaluated as it provides the most flexibility as well as
deterministic high availability for on-premise database clouds. Please refer to the Appendix for additional
information resources.
The next high level step is to determine the service groupings and base sizing. This will involve answering such
questions as which services need to run on the same servers, which must be exclusive, or which must be
dispersed as well as services that are required to be singletons.
When sizing a on-premise database cloud or DBaaS deployment, the tendency is to make use of multi-threaded
CPU cores in order to increase the effective number of CPUs that each database sees in the hope that more
databases can be hosted per node. The curves in Figure 3 should be observed as a warning that the level of
requests per CPU is significantly reduced before response time goes to infinity and the system is in overload. It
should also be noted that predicable performance is no longer achievable because the OS scheduler is now
directing database workload scheduling and not the database’s resource manager. This results in the CPU cost
per database call rising with utilization instead of staying constant as with a single threaded core.
4
Figure 3: Consolidation Management Problem
Perhaps the most challenging step is to establish the business criticality of each service, answering such
questions as:
» Which are the services that need to be online first?
» Which are the services that need to be the last standing?
» Which services can I borrow resources from should a workload surge occur?
» Which services can I shut down should a surge or failure occur?
Fortunately, these questions don’t have to have static answers if a policy-managed deployment type is selected,
as different business priorities can be expressed in different policies that can be switched in when appropriate. At
the same time, legacy databases can coexist within their fenced servers within the Generic server pool yet still be
fully supported.
Finally, services need to be group or “classified” into those that need to be tracked for performance and those
that simply need to be measured. This classification may be performed by using the QoS Management Policy
Editor integrated into Enterprise Manager Cloud Control to create user-defined labels or tags that group
workloads for both measurement and assigning performance objectives that can be monitored or managed to as
will be described in later phases.
Figure 4 shows where the QoS Management functionality can be found in Enterprise Manager Cloud Control.
Note that it is accessed from the Cluster target Administration menu. This is because the scope of management
is currently the entire cluster of RAC databases.
5
Figure 4: Creating a QoS Management Policy Set
QoS Management generates a default policy set by discovering the entire set of cluster-managed database
services currently registered and creating a performance class for each one. This can be seen in Figure 5. Each
Performance Class has one or more Classifiers which are the Boolean set expressions shown in the figure.
Figure 5: Overview of Performance Classes and Classifiers
In some cases, there may be significantly different types of workloads using the same service. Under this
condition additional performance classes may be created that can differentiate the workloads using the database
6
session parameters if Module, Action UserName and Program. An example of differentiating browsing users of
the sales service from those who are purchasing is seen in Figure 6 where salescart_pc Performance Class is
being created specifying a different database user. Session Module, Action and Program can be populated in the
same way.
Figure 6: Creating a Performance Class
Should a group of services have similar resource use and performance objectives where it is desired to manage
them together, this can be done by adding additional classifiers to a single performance class.
Phase 2: Runtime Measure the Deployment
Once the planning phase is completed, the measurement phase may begin. This is not the same type of
measurement that occurs in single application Q/A or testing but in either the production environment or a test
one where all databases and these services are running as they would in production. To set up the ability to
perform these actual runtime measurements, a measure-only Performance Policy is created in the same QoS
Management Policy Editor. This is shown in Figure 7. What distinguishes this policy from others is that no
performance objectives are specified and the Measure Only box for each performance class is checked.
7
Figure 7: Creating a QoS Management - Measure-Only Policy
Once the Policy Editor wizard is completed and the policy set submitted to the QoS Management server with this
measure-only policy activated, the QoS Management Dashboard is displayed as seen in Figure 8. Note that all of
the performance classes are listed and the actual server pools where work is occurring are specified.
Figure 8: QoS Management Runtime Measure Dashboard
8
Examining the displayed bar graph for each performance class in Figure 9, two important metrics are displayed.
The blue bar shows the actual portion of the response time that represents the use of system resources such as
CPU memory and I/O. When hovered over the value displayed in seconds represents the absolute best
performance that can be achieved with the deployed resource capability. The gray bar shows the actual portion
of the response time that represents the wait for system resources. This time is a function of how busy the
system resources are and may be altered via runtime resource management controls. When added together the
two represent the actual performance which would be the minimum recommended performance objective set for
this performance class given the other workloads.
Figure 9: Runtime Measurement Detail
Moving down the QoS Management Dashboard is a table that breaks down the resource wait time for each
performance class into four categories as shown in Figure 10 – CPU, Global Cache, IO, and Other.
Figure 10: Resource Wait Detail by Performance Class
These metrics are very useful in understanding whether there are runtime issues beyond simple resource
availability with a workload. For example, if Global Cache wait time was the largest and thus the bottleneck, it is
most likely that the workload doesn’t scale well across more than one instance and its service should be a
singleton or the data should be partitioned. If Other wait is the bottleneck, this means that there are SQL issues
in the database that should be investigated via an AWR report.
Once this phase is run during all the different workload periods, the metrics will provide a baseline set of
minimum performance objective values that may be used in the next phase. It will also provide data that will help
determine if base sizing and resources are sufficient to meet the business objectives and whether multiple
policies may be useful in meeting these.
9
Phase 3: Runtime Monitor the Deployment
The third phase is to actually monitor the deployment using performance objectives derived from the previous
phase. This requires a different QoS Management performance policy which can be added to the policy set using
the Policy Editor in EM Cloud Control. Figure 11 shows an example of such a policy. Note that this monitor policy
is quite similar to the previous measure-only policy. The difference is that now actual performance objective
values are entered.
Figure 11: QoS Management - Monitor Policy
Once this policy is submitted and activated, the QoS Management Dashboard changes as displayed in Figure
12. Additional colored bars appear and the Performance Satisfaction Metric column also becomes relevant.
10
Figure 12: QoS Management Monitor Dashboard
Figure 13 provides a close look at the right-most two columns. The Resource Use vs Wait Time column now has
additional bars. Since there is now a performance objective specified in the policy that is represented by the right
end of the entire bar. As the response time is the sum of the blue and gray bars, the green bar represents the
additional time until the performance objective is met. This is known as the headroom and can be viewed as the
amount of shareable resources that can be contributed without exceeding the performance objective. This will be
clear in the next phase.
Figure 13: Monitor Measurement Detail
If a performance objective is specified too optimistically and is exceeded the gray bar becomes red to indicate
that the wait for resources has caused the response tome of the associated performance class to exceed its
performance objective. To provide an indicator of how far it has exceeded, the performance objective is
displayed as a blue line with the red bar.
The Performance Satisfaction Metric (PSM) is a unique QoS Management metric as it quantifies in a normalized
way how the performance class is doing against its objective. Whether a performance objective is 5ms or 5s, it
11
reports a value between -100% and +100% to indicate the degree response time is meeting or violating its
objective. The PSM column is a binary indication as to how the performance classes have been doing against
their objectives for the last 5 minutes thereby providing trending information. In this example, the continuous red
indicates that the performance objective should be re-evaluated if the load is in expected normal range.
It is obviously not convenient or efficient to require constant monitoring of the QoS Management Dashboard;
therefore, support is provided in the EMCC notification system for reporting negative PSMs that persist for userspecified times. Both warning and critical levels can be alerted based upon specified durations for each
performance class as shown in Figure 14.
Figure 14: EM Negative PSM Duration Notification Setup
This phase will likely need to be executed in an iterative fashion to establish realistic performance objectives.
During this process, it may be determined that a single policy is not sufficient to capture the workload phases
such as daytime, nighttime, weekends, end of quarter, etc. In that case create multiple policies which can be
either switched manually or via a scheduler such as EMCC or CRON and the included qosctl command line
utility.
Once this phase has been completed, the decision can be made to move to the final phase. If, for example there
are sufficient resources under all demand phases, it may not be necessary to transition to the runtime
management phase. However, if this on-premise database cloud is exposed to open workloads such as the
Internet, the ability to respond with just-in-time intelligent resource allocation may be critical to maintain business
continuity.
Phase 4: Runtime Manage the Deployment
This final phase brings resource agility into runtime management of a on-premise database cloud. Many
resource management systems are, in the end, simply issue-response sets of thresholds and rules. While they
may work for simple systems, they are inadequate for the complexity of an enterprise database cloud as
resources cannot be provisioned on the fly. Instead resource trade-offs and agility within the existing deployment
must be able to be evaluated, taking into account business priorities. This is where the expert system in the QoS
Management server comes into play.
As with the other phases, this one requires a different policy. Figure 15 is an example of one such policy. It is
differentiated from the previous ones in several important ways. First, it takes into account the rank of each
performance class which is settable to one of 5 levels. This rank expresses how critical to the business it is for a
performance class to continue to meet its objective.
12
Figure 15: QoS Management – Manage Policy
Second, all performance classes that are to be managed have their Measure Only checkmarks removed. Any
that remain checked will be considered donors should resources be required by those classes that are managed.
Third, a list of resource management type actions is offered to authorize QoS Management to autonomously that
that action should it be required. This option is not recommended to be enabled until the accuracy and
effectiveness of the recommendations and actions have been in production for some time.
Finally, there is the Server Pool Override Directive section which is used when multiple policies require different
known base resource allocations. An example would be adding a server to a pool responsible for end of quarter
reporting.
Once this type of policy is active, instead of simply being alerted to a performance class problem and viewing the
extent of the issue as in the monitor phase, the QoS Management Dashboard displays a recommended action
which will have a positive effect on relieving the bottleneck by trading off resources between workloads. Figure
16 shows an example of one such recommendation. In this case the performance class, salescart_pc, is
experiencing a resource bottleneck in trying to get access to the CPU. The QoS Management performance
model evaluation, determined that moving 15 CPU shares from the prod HCMPDB database to the SALESPDB
database will provide more CPU time to the more critical workload and reduce its bottleneck thereby improving
response time.
13
Figure 16: QoS Management Recommendation
If greater insight is desired before clicking the Implement button, full details on the recommendation are available
as shown in Figure 17. Since this is a trade-off evaluation a full view of the projected positive and negative
impacts to each performance class is presented along with the improvement to the target class.
Figure 17: Recommendation Details and Performance Projections
QoS Management acts as a governor on existing Oracle resource management functionality. It can adjust CPU
shares within a single database to manage schema consolidated services, the number of CPU shares allocated
to each PDB in a CDB multitenant database as in the example, the movement of CPUs between databases
permitting management of multiple databases sharing the same servers and finally in policy-managed
deployments, the ability to move servers between server pools allows for cluster consolidation.
When altering resource allocations, it’s critically important to track the performance over time. Either singularly or
overall, various graphical metric views can be access from the QoS Management Dashboard. Figure 18 is an
overall view of the demand that the cluster is seeing and the obvious surge to the salescart_pc that caused it to
violate its objective.
Figure 18: Overall Demand per Performance Class
Figure 19 shows a companion graph of the PSM values during the same period and how the demand surge
impacted multiple performance classes. It also shows how through multiple recommendations and reallocations,
performance was ultimately restored.
14
Figure 19: Overall PSM Status per Performance Class
Currently this management phase is targeted at managing workloads whose demand is independent of the
response time. These open workloads are the dominant type on the Internet and are particularly difficult to
provision for. By clustering resources and making them agile, workload surges can be efficiently accommodated
thereby minimizing idle capacity.
Conclusion
The desire for database consolidation and a provision-on-demand DBaaS to meet the growing demand without
the costs of growing datacenters brings with it new infrastructure functionality and a management paradigm that
is both flexible in its resource allocation and deterministic in its operational and failure behavior. There is a
learning curve with this type of change, but the ability to implement these types of deployments in managed best
practice phases mitigates risk while delivering greater resource utilization and subsequent higher return on
investment.
Figure 20: QoS Management in Action
15
Through its continuous measuring, monitoring and managing of the Oracle RAC on-premise database cloud or
DBaaS deployments as illustrated in Figure 20, Oracle Database Quality of Service Management delivers the
following key elements for runtime management:
» Cluster-wide real-time dashboard view of all database workloads
» Continuous workload health view
» Real-time resource bottleneck identification
» Workload-specific notifications of performance issues
» Intelligent and targeted bottleneck resolution recommendations
» Action audit trail and performance history
Appendix
Further information on Policy-Managed RAC databases, Clusterware policies and server pools as well as QoS
Management is available from the following links:
Oracle QoS Management 12c Documentation
Oracle Autonomous Health Framework 12c Documentation
Oracle QoS Management on OTN
Oracle QoS Management FAQ
Oracle Database 12c: Why and How You Should Be Using Policy-Managed Oracle RAC Databases
16
Oracle Corporation, World Headquarters
Worldwide Inquiries
500 Oracle Parkway
Phone: +1.650.506.7000
Redwood Shores, CA 94065, USA
Fax: +1.650.506.7200
CONNE CT WI TH US
blogs.oracle.com/oracle
facebook.com/oracle
twitter.com/oracle
oracle.com
Copyright © 2016, Oracle and/or its affiliates. All rights reserved. This document is provided for information purposes only, and the
contents hereof are subject to change without notice. This document is not warranted to be error-free, nor subject to any other
warranties or conditions, whether expressed orally or implied in law, including implied warranties and conditions of merchantability or
fitness for a particular purpose. We specifically disclaim any liability with respect to this document, and no contractual obligations are
formed either directly or indirectly by this document. This document may not be reproduced or transmitted in any form or by any means,
electronic or mechanical, for any purpose, without our prior written permission.
Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners.
Intel and Intel Xeon are trademarks or registered trademarks of Intel Corporation. All SPARC trademarks are used under license and
are trademarks or registered trademarks of SPARC International, Inc. AMD, Opteron, the AMD logo, and the AMD Opteron logo are
trademarks or registered trademarks of Advanced Micro Devices. UNIX is a registered trademark of The Open Group. 0116
Oracle Database 12c: Quality of Service Management - Monitoring and Managing Oracle RAC Database Performance
March 2017
Author: Mark Scardina, Oracle Corporation
2
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement