openstack-manuals/doc/arch-design/compute_focus/section_tech_considerations_compute_focus.xml

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE section [
<!ENTITY % openstack SYSTEM "../../common/entities/openstack.ent">
%openstack;
]>
<section xmlns="http://docbook.org/ns/docbook"
  xmlns:xi="http://www.w3.org/2001/XInclude"
  xmlns:xlink="http://www.w3.org/1999/xlink"
  version="5.0"
  xml:id="technical-considerations-compute-focus">
    <?dbhtml stop-chunking?>
    <title>Technical considerations</title>
    <para>In a compute-focused OpenStack cloud, the type of instance
        workloads being provisioned heavily influences technical
        decision making. For example, specific use cases that demand
        multiple short running jobs present different requirements
        than those that specify long-running jobs, even though both
        situations are considered "compute focused."</para>
    <para>Public and private clouds require deterministic capacity
        planning to support elastic growth in order to meet user SLA
        expectations. Deterministic capacity planning is the path to
        predicting the effort and expense of making a given process
        consistently performant. This process is important because,
        when a service becomes a critical part of a user's
        infrastructure, the user's fate becomes wedded to the SLAs of
        the cloud itself. In cloud computing, a service’s performance
        will not be measured by its average speed but rather by the
        consistency of its speed.</para>
    <para>There are two aspects of capacity planning to consider:
        planning the initial deployment footprint, and planning
        expansion of it to stay ahead of the demands of cloud
        users.</para>
    <para>Planning the initial footprint for an OpenStack deployment
        is typically done based on existing infrastructure workloads
        and estimates based on expected uptake.</para>
    <para>The starting point is the core count of the cloud. By
        applying relevant ratios, the user can gather information
        about:</para>
    <itemizedlist>
        <listitem>
            <para>The number of instances expected to be available
                concurrently: (overcommit fraction × cores) / virtual
                cores per instance</para>
        </listitem>
        <listitem>
            <para>How much storage is required: flavor disk size ×
                number of instances</para>
        </listitem>
    </itemizedlist>
    <para>These ratios can be used to determine the amount of
        additional infrastructure needed to support the cloud. For
        example, consider a situation in which you require 1600
        instances, each with 2 vCPU and 50 GB of storage. Assuming the
        default overcommit rate of 16:1, working out the math provides
        an equation of:</para>
    <itemizedlist>
        <listitem>
            <para>1600 = (16 &times; (number of physical cores)) / 2</para>
        </listitem>
        <listitem>
            <para>storage required = 50&nbsp;GB &times; 1600</para>
        </listitem>
    </itemizedlist>
    <para>On the surface, the equations reveal the need for 200
        physical cores and 80&nbsp;TB of storage for
        <filename>/var/lib/nova/instances/</filename>. However,
        it is also important to
        look at patterns of usage to estimate the load that the API
        services, database servers, and queue servers are likely to
        encounter.</para>
    <para>Consider, for example, the differences between a cloud that
        supports a managed web-hosting platform with one running
        integration tests for a development project that creates one
        instance per code commit. In the former, the heavy work of
        creating an instance happens only every few months, whereas
        the latter puts constant heavy load on the cloud controller.
        The average instance lifetime must be considered, as a larger
        number generally means less load on the cloud
        controller.</para>
    <para>Aside from the creation and termination of instances, the
        impact of users must be considered when accessing the service,
        particularly on nova-api and its associated database. Listing
        instances garners a great deal of information and, given the
        frequency with which users run this operation, a cloud with a
        large number of users can increase the load significantly.
        This can even occur unintentionally. For example, the
        OpenStack Dashboard instances tab refreshes the list of
        instances every 30 seconds, so leaving it open in a browser
        window can cause unexpected load.</para>
    <para>Consideration of these factors can help determine how many
        cloud controller cores are required. A server with 8 CPU cores
        and 8 GB of RAM server would be sufficient for up to a rack of
        compute nodes, given the above caveats.</para>
    <para>Key hardware specifications are also crucial to the
        performance of user instances. Be sure to consider budget and
        performance needs, including storage performance
        (spindles/core), memory availability (RAM/core), network
        bandwidth (Gbps/core), and overall CPU performance
        (CPU/core).</para>
    <para>The cloud resource calculator is a useful tool in examining
        the impacts of different hardware and instance load outs. It
        is available at: <link xlink:href="https://github.com/noslzzp/cloud-resource-calculator/blob/master/cloud-resource-calculator.ods">https://github.com/noslzzp/cloud-resource-calculator/blob/master/cloud-resource-calculator.ods</link>
    </para>
    <section xml:id="expansion-planning-compute-focus">
        <title>Expansion planning</title>
    <para>A key challenge faced when planning the expansion of cloud
        compute services is the elastic nature of cloud infrastructure
        demands. Previously, new users or customers would be forced to
        plan for and request the infrastructure they required ahead of
        time, allowing time for reactive procurement processes. Cloud
        computing users have come to expect the agility provided by
        having instant access to new resources as they are required.
        Consequently, this means planning should be delivered for
        typical usage, but also more importantly, for sudden bursts in
        usage.</para>
    <para>Planning for expansion can be a delicate balancing act.
        Planning too conservatively can lead to unexpected
        oversubscription of the cloud and dissatisfied users. Planning
        for cloud expansion too aggressively can lead to unexpected
        underutilization of the cloud and funds spent on operating
        infrastructure that is not being used efficiently.</para>
    <para>The key is to carefully monitor the spikes and valleys in
        cloud usage over time. The intent is to measure the
        consistency with which services can be delivered, not the
        average speed or capacity of the cloud. Using this information
        to model performance results in capacity enables users to more
        accurately determine the current and future capacity of the
        cloud.</para></section>
    <section xml:id="cpu-and-ram-compute-focus"><title>CPU and RAM</title>
    <para>(Adapted from:
        <link xlink:href="http://docs.openstack.org/openstack-ops/content/compute_nodes.html#cpu_choice">http://docs.openstack.org/openstack-ops/content/compute_nodes.html#cpu_choice</link>)</para>
    <para>In current generations, CPUs have up to 12 cores. If an
        Intel CPU supports Hyper-Threading, those 12 cores are doubled
        to 24 cores. If a server is purchased that supports multiple
        CPUs, the number of cores is further multiplied.
        Hyper-Threading is Intel's proprietary simultaneous
        multi-threading implementation, used to improve
        parallelization on their CPUs. Consider enabling
        Hyper-Threading to improve the performance of multithreaded
        applications.</para>
    <para>Whether the user should enable Hyper-Threading on a CPU
        depends upon the use case. For example, disabling
        Hyper-Threading can be beneficial in intense computing
        environments. Performance testing conducted by running local
        workloads with both Hyper-Threading on and off can help
        determine what is more appropriate in any particular
        case.</para>
    <para>If the Libvirt/KVM hypervisor driver are the intended use
        cases, then the CPUs used in the compute nodes must support
        virtualization by way of the VT-x extensions for Intel chips
        and AMD-v extensions for AMD chips to provide full
        performance.</para>
    <para>OpenStack enables the user to overcommit CPU and RAM on
        compute nodes. This allows an increase in the number of
        instances running on the cloud at the cost of reducing the
        performance of the instances. OpenStack Compute uses the
        following ratios by default:</para>
    <itemizedlist>
        <listitem>
            <para>CPU allocation ratio: 16:1</para>
        </listitem>
        <listitem>
            <para>RAM allocation ratio: 1.5:1</para>
        </listitem>
    </itemizedlist>
    <para>The default CPU allocation ratio of 16:1 means that the
        scheduler allocates up to 16 virtual cores per physical core.
        For example, if a physical node has 12 cores, the scheduler
        sees 192 available virtual cores. With typical flavor
        definitions of 4 virtual cores per instance, this ratio would
        provide 48 instances on a physical node.</para>
    <para>Similarly, the default RAM allocation ratio of 1.5:1 means
        that the scheduler allocates instances to a physical node as
        long as the total amount of RAM associated with the instances
        is less than 1.5 times the amount of RAM available on the
        physical node.</para>
    <para>For example, if a physical node has 48 GB of RAM, the
        scheduler allocates instances to that node until the sum of
        the RAM associated with the instances reaches 72 GB (such as
        nine instances, in the case where each instance has 8 GB of
        RAM).</para>
    <para>The appropriate CPU and RAM allocation ratio must be
        selected based on particular use cases.</para></section>
    <section xml:id="additional-hardware-compute-focus">
      <title>Additional hardware</title>
    <para>Certain use cases may benefit from exposure to additional
        devices on the compute node. Examples might include:</para>
    <itemizedlist>
        <listitem>
            <para>High performance computing jobs that benefit from
                the availability of graphics processing units (GPUs)
                for general-purpose computing.</para>
        </listitem>
    </itemizedlist>
    <itemizedlist>
        <listitem>
            <para>Cryptographic routines that benefit from the
                availability of hardware random number generators to
                avoid entropy starvation.</para>
        </listitem>
        <listitem>
            <para>Database management systems that benefit from the
                availability of SSDs for ephemeral storage to maximize
                read/write time when it is required.</para>
        </listitem>
    </itemizedlist>
    <para>Host aggregates are used to group hosts that share similar
        characteristics, which can include hardware similarities. The
        addition of specialized hardware to a cloud deployment is
        likely to add to the cost of each node, so careful
        consideration must be given to whether all compute nodes, or
        just a subset which is targetable using flavors, need the
        additional customization to support the desired
        workloads.</para></section>
    <section xml:id="utilization"><title>Utilization</title>
    <para>Infrastructure-as-a-Service offerings, including OpenStack,
        use flavors to provide standardized views of virtual machine
        resource requirements that simplify the problem of scheduling
        instances while making the best use of the available physical
        resources.</para>
    <para>In order to facilitate packing of virtual machines onto
        physical hosts, the default selection of flavors are
        constructed so that the second largest flavor is half the size
        of the largest flavor in every dimension. It has half the
        vCPUs, half the vRAM, and half the ephemeral disk space. The
        next largest flavor is half that size again. As a result,
        packing a server for general purpose computing might look
        conceptually something like this figure:
    <mediaobject>
        <imageobject>
            <imagedata contentwidth="4in"
                fileref="../images/Compute_Tech_Bin_Packing_General1.png"
            />
        </imageobject>
    </mediaobject></para>
    <para>On the other hand, a CPU optimized packed server might look
        like the following figure:
    <mediaobject>
        <imageobject>
            <imagedata contentwidth="4in"
                fileref="../images/Compute_Tech_Bin_Packing_CPU_optimized1.png"
            />
        </imageobject>
    </mediaobject></para>
    <para>These default flavors are well suited to typical load outs
        for commodity server hardware. To maximize utilization,
        however, it may be necessary to customize the flavors or
        create new ones, to better align instance sizes to the
        available hardware.</para>
    <para>Workload characteristics may also influence hardware choices
        and flavor configuration, particularly where they present
        different ratios of CPU versus RAM versus HDD
        requirements.</para>
    <para>For more information on Flavors refer to:
        <link xlink:href="http://docs.openstack.org/openstack-ops/content/flavors.html">http://docs.openstack.org/openstack-ops/content/flavors.html</link></para>
    </section>
    <section xml:id="performance-compute-focus"><title>Performance</title>
    <para>The infrastructure of a cloud should not be shared, so that
        it is possible for the workloads to consume as many resources
        as are made available, and accommodations should be made to
        provide large scale workloads.</para>
    <para>The duration of batch processing differs depending on
        individual workloads that are launched. Time limits range from
        seconds, minutes to hours, and as a result it is considered
        difficult to predict when resources will be used, for how
        long, and even which resources will be used.</para>
    </section>
    <section xml:id="security-compute-focus"><title>Security</title>
    <para>The security considerations needed for this scenario are
        similar to those of the other scenarios discussed in this
        book.</para>
    <para>A security domain comprises users, applications, servers
        or networks that share common trust requirements and
        expectations within a system. Typically they have the same
        authentication and authorization requirements and
        users.</para>
    <para>These security domains are:</para>
    <orderedlist>
        <listitem>
            <para>Public</para>
        </listitem>
        <listitem>
            <para>Guest</para>
        </listitem>
        <listitem>
            <para>Management</para>
        </listitem>
        <listitem>
            <para>Data</para>
        </listitem>
    </orderedlist>
    <para>These security domains can be mapped individually to the
        installation, or they can also be combined. For example, some
        deployment topologies combine both guest and data domains onto
        one physical network, whereas in other cases these networks
        are physically separated. In each case, the cloud operator
        should be aware of the appropriate security concerns. Security
        domains should be mapped out against specific OpenStack
        deployment topology. The domains and their trust requirements
        depend upon whether the cloud instance is public, private, or
        hybrid.</para>
    <para>The public security domain is an entirely untrusted area of
        the cloud infrastructure. It can refer to the Internet as a
        whole or simply to networks over which the user has no
        authority. This domain should always be considered
        untrusted.</para>
    <para>Typically used for compute instance-to-instance traffic, the
        guest security domain handles compute data generated by
        instances on the cloud; not services that support the
        operation of the cloud, for example API calls. Public cloud
        providers and private cloud providers who do not have
        stringent controls on instance use or who allow unrestricted
        Internet access to instances should consider this domain to be
        untrusted. Private cloud providers may want to consider this
        network as internal and therefore trusted only if they have
        controls in place to assert that they trust instances and all
        their tenants.</para>
    <para>The management security domain is where services interact.
        Sometimes referred to as the "control plane", the networks in
        this domain transport confidential data such as configuration
        parameters, user names, and passwords. In most deployments this
        domain is considered trusted.</para>
    <para>The data security domain is concerned primarily with
        information pertaining to the storage services within
        OpenStack. Much of the data that crosses this network has high
        integrity and confidentiality requirements and depending on
        the type of deployment there may also be strong availability
        requirements. The trust level of this network is heavily
        dependent on deployment decisions and as such we do not assign
        this any default level of trust.</para>
    <para>When deploying OpenStack in an enterprise as a private cloud
        it is assumed to be behind a firewall and within the trusted
        network alongside existing systems. Users of the cloud are
        typically employees or trusted individuals that are bound by
        the security requirements set forth by the company. This tends
        to push most of the security domains towards a more trusted
        model. However, when deploying OpenStack in a public-facing
        role, no assumptions can be made and the attack vectors
        significantly increase. For example, the API endpoints and the
        software behind it will be vulnerable to potentially hostile
        entities wanting to gain unauthorized access or prevent access
        to services. This can result in loss of reputation and must be
        protected against through auditing and appropriate
        filtering.</para>
    <para>Consideration must be taken when managing the users of the
        system, whether it is the operation of public or private
        clouds. The identity service allows for LDAP to be part of the
        authentication process, and includes such systems as an
        OpenStack deployment that may ease user management if
        integrated into existing systems.</para>
    <para>It is strongly recommended that the API services are placed
        behind hardware that performs SSL termination. API services
        transmit user names, passwords, and generated tokens between
        client machines and API endpoints and therefore must be
        secured.</para>
    <para>More information on OpenStack Security can be found
        at <link xlink:href="http://docs.openstack.org/security-guide/">http://docs.openstack.org/security-guide/</link></para>
    </section>
    <section xml:id="openstack-components-compute-focus">
      <title>OpenStack components</title>
    <para>Due to the nature of the workloads that will be used in this
        scenario, a number of components will be highly beneficial in
        a Compute-focused cloud. This includes the typical OpenStack
        components:</para>
    <itemizedlist>
        <listitem>
            <para>OpenStack Compute (nova)</para>
        </listitem>
        <listitem>
            <para>OpenStack Image Service (glance)</para>
        </listitem>
        <listitem>
            <para>OpenStack Identity (keystone)</para>
        </listitem>
    </itemizedlist>
    <para>Also consider several specialized components:</para>
    <itemizedlist>
        <listitem>
            <para><glossterm>Orchestration</glossterm> module
            (<glossterm>heat</glossterm>)</para>
        </listitem>
    </itemizedlist>
    <para>It is safe to assume that, given the nature of the
        applications involved in this scenario, these will be heavily
        automated deployments. Making use of Orchestration will be highly
        beneficial in this case. Deploying a batch of instances and
        running an automated set of tests can be scripted, however it
        makes sense to use the Orchestration module
        to handle all these actions.</para>
    <itemizedlist>
        <listitem>
            <para>Telemetry module (ceilometer)</para>
        </listitem>
    </itemizedlist>
    <para>Telemetry and the alarms it generates are required
        to support autoscaling of instances using
        Orchestration. Users that are not using the
        Orchestration module do not need to deploy the Telemetry module and
        may choose to use other external solutions to fulfill their
        metering and monitoring requirements.</para>
    <para>See also:
        <link xlink:href="http://docs.openstack.org/openstack-ops/content/logging_monitoring.html">http://docs.openstack.org/openstack-ops/content/logging_monitoring.html</link></para>
    <itemizedlist>
        <listitem>
            <para>OpenStack Block Storage (cinder)</para>
        </listitem>
    </itemizedlist>
    <para>Due to the burst-able nature of the workloads and the
        applications and instances that will be used for batch
        processing, this cloud will utilize mainly memory or CPU, so
        the need for add-on storage to each instance is not a likely
        requirement. This does not mean that OpenStack Block Storage
        (cinder) will not be used in the infrastructure, but
        typically it will not be used as a central component.</para>
    <itemizedlist>
        <listitem>
            <para>Networking</para>
        </listitem>
    </itemizedlist>
    <para>When choosing a networking platform, ensure that it either
        works with all desired hypervisor and container technologies
        and their OpenStack drivers, or includes an implementation of
        an ML2 mechanism driver. Networking platforms that provide ML2
        mechanisms drivers can be mixed.</para></section>
</section>