openstack-manuals/doc/ops-guide/source/arch_scaling.rst

Scaling

Whereas traditional applications required larger hardware to scale ("vertical scaling"), cloud-based applications typically request more, discrete hardware ("horizontal scaling"). If your cloud is successful, eventually you must add resources to meet the increasing demand.

To suit the cloud paradigm, OpenStack itself is designed to be horizontally scalable. Rather than switching to larger servers, you procure more servers and simply install identically configured services. Ideally, you scale out and load balance among groups of functionally identical services (for example, compute nodes or nova-api nodes), that communicate on a message bus.

The Starting Point

Determining the scalability of your cloud and how to improve it is an exercise with many variables to balance. No one solution meets everyone's scalability goals. However, it is helpful to track a number of metrics. Since you can define virtual hardware templates, called "flavors" in OpenStack, you can start to make scaling decisions based on the flavors you'll provide. These templates define sizes for memory in RAM, root disk size, amount of ephemeral data disk space available, and number of cores for starters.

The default OpenStack flavors are shown in table_default_flavors.

Table. OpenStack default flavors

Name	Virtual cores	Memory	Disk	Ephemeral
m1.tiny	1	512 MB	1 GB	0 GB
m1.small	1	2 GB	10 GB	20 GB
m1.medium	2	4 GB	10 GB	40 GB
m1.large	4	8 GB	10 GB	80 GB
m1.xlarge	8	16 GB	10 GB	160 GB

The starting point for most is the core count of your cloud. By applying some ratios, you can gather information about:

• The number of virtual machines (VMs) you expect to run, ((overcommit fraction × cores) / virtual cores per instance)
• How much storage is required (flavor disk size × number of instances)

You can use these ratios to determine how much additional infrastructure you need to support your cloud.

Here is an example using the ratios for gathering scalability information for the number of VMs expected as well as the storage needed. The following numbers support (200 / 2) × 16 = 1600 VM instances and require 80 TB of storage for /var/lib/nova/instances:

• 200 physical cores.
• Most instances are size m1.medium (two virtual cores, 50 GB of storage).
• Default CPU overcommit ratio (cpu_allocation_ratio in nova.conf) of 16:1.

Note

Regardless of the overcommit ratio, an instance can not be placed on any physical node with fewer raw (pre-overcommit) resources than instance flavor requires.

However, you need more than the core count alone to estimate the load that the API services, database servers, and queue servers are likely to encounter. You must also consider the usage patterns of your cloud.

As a specific example, compare a cloud that supports a managed web-hosting platform with one running integration tests for a development project that creates one VM per code commit. In the former, the heavy work of creating a VM happens only every few months, whereas the latter puts constant heavy load on the cloud controller. You must consider your average VM lifetime, as a larger number generally means less load on the cloud controller.

Aside from the creation and termination of VMs, you must consider the impact of users 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 occur even without their knowledge—leaving the OpenStack dashboard instances tab open in the browser refreshes the list of VMs every 30 seconds.

After you consider these factors, you can determine how many cloud controller cores you require. A typical eight core, 8 GB of RAM server is sufficient for up to a rack of compute nodes — given the above caveats.

You must also consider key hardware specifications for the performance of user VMs, as well as budget and performance needs, including storage performance (spindles/core), memory availability (RAM/core), network bandwidth (Gbps/core), and overall CPU performance (CPU/core).

Tip

For a discussion of metric tracking, including how to extract metrics from your cloud, see ops_logging_monitoring.

Adding Cloud Controller Nodes

You can facilitate the horizontal expansion of your cloud by adding nodes. Adding compute nodes is straightforward—they are easily picked up by the existing installation. However, you must consider some important points when you design your cluster to be highly available.

Recall that a cloud controller node runs several different services. You can install services that communicate only using the message queue internally—nova-scheduler and nova-console—on a new server for expansion. However, other integral parts require more care.

You should load balance user-facing services such as dashboard, nova-api, or the Object Storage proxy. Use any standard HTTP load-balancing method (DNS round robin, hardware load balancer, or software such as Pound or HAProxy). One caveat with dashboard is the VNC proxy, which uses the WebSocket protocol—something that an L7 load balancer might struggle with. See also Horizon session storage.

You can configure some services, such as nova-api and glance-api, to use multiple processes by changing a flag in their configuration file—allowing them to share work between multiple cores on the one machine.

Tip

Several options are available for MySQL load balancing, and the supported AMQP brokers have built-in clustering support. Information on how to configure these and many of the other services can be found in operations.

Segregating Your Cloud

When you want to offer users different regions to provide legal considerations for data storage, redundancy across earthquake fault lines, or for low-latency API calls, you segregate your cloud. Use one of the following OpenStack methods to segregate your cloud: cells, regions, availability zones, or host aggregates.

Each method provides different functionality and can be best divided into two groups:

• Cells and regions, which segregate an entire cloud and result in running separate Compute deployments.
• Availability zones <availability zone> and host aggregates, which merely divide a single Compute deployment.

table_segregation_methods provides a comparison view of each segregation method currently provided by OpenStack Compute.

Table. OpenStack segregation methods

	Cells	Regions	Availability zones	Host aggregates
Use when you need	A single API endpoint for compute, or you require a second level of scheduling.	Discrete regions with separate API endpoints and no coordination between regions.	Logical separation within your nova deployment for physical isolation or redundancy.	To schedule a group of hosts with common features.
Example	A cloud with multiple sites where you can schedule VMs "anywhere" or on a particular site.	A cloud with multiple sites, where you schedule VMs to a particular site and you want a shared infrastructure.	A single-site cloud with equipment fed by separate power supplies.	Scheduling to hosts with trusted hardware support.
Overhead	Considered experimental. A new service, nova-cells. Each cell has a full nova installation except nova-api.	A different API endpoint for every region. Each region has a full nova installation.	Configuration changes to nova.conf.	Configuration changes to nova.conf.
Shared services	Keystone, nova-api	Keystone	Keystone, All nova services	Keystone, All nova services

Cells and Regions

OpenStack Compute cells are designed to allow running the cloud in a distributed fashion without having to use more complicated technologies, or be invasive to existing nova installations. Hosts in a cloud are partitioned into groups called cells. Cells are configured in a tree. The top-level cell ("API cell") has a host that runs the nova-api service, but no nova-compute services. Each child cell runs all of the other typical nova-* services found in a regular installation, except for the nova-api service. Each cell has its own message queue and database service and also runs nova-cells, which manages the communication between the API cell and child cells.

This allows for a single API server being used to control access to multiple cloud installations. Introducing a second level of scheduling (the cell selection), in addition to the regular nova-scheduler selection of hosts, provides greater flexibility to control where virtual machines are run.

Unlike having a single API endpoint, regions have a separate API endpoint per installation, allowing for a more discrete separation. Users wanting to run instances across sites have to explicitly select a region. However, the additional complexity of a running a new service is not required.

The OpenStack dashboard (horizon) can be configured to use multiple regions. This can be configured through the AVAILABLE_REGIONS parameter.

Availability Zones and Host Aggregates

You can use availability zones, host aggregates, or both to partition a nova deployment.

Availability zones are implemented through and configured in a similar way to host aggregates.

However, you use them for different reasons.

Availability zone

This enables you to arrange OpenStack compute hosts into logical groups and provides a form of physical isolation and redundancy from other availability zones, such as by using a separate power supply or network equipment.

You define the availability zone in which a specified compute host resides locally on each server. An availability zone is commonly used to identify a set of servers that have a common attribute. For instance, if some of the racks in your data center are on a separate power source, you can put servers in those racks in their own availability zone. Availability zones can also help separate different classes of hardware.

When users provision resources, they can specify from which availability zone they want their instance to be built. This allows cloud consumers to ensure that their application resources are spread across disparate machines to achieve high availability in the event of hardware failure.

Host aggregates zone

This enables you to partition OpenStack Compute deployments into logical groups for load balancing and instance distribution. You can use host aggregates to further partition an availability zone. For example, you might use host aggregates to partition an availability zone into groups of hosts that either share common resources, such as storage and network, or have a special property, such as trusted computing hardware.

A common use of host aggregates is to provide information for use with the nova-scheduler. For example, you might use a host aggregate to group a set of hosts that share specific flavors or images.

The general case for this is setting key-value pairs in the aggregate metadata and matching key-value pairs in flavor's extra_specs metadata. The AggregateInstanceExtraSpecsFilter in the filter scheduler will enforce that instances be scheduled only on hosts in aggregates that define the same key to the same value.

An advanced use of this general concept allows different flavor types to run with different CPU and RAM allocation ratios so that high-intensity computing loads and low-intensity development and testing systems can share the same cloud without either starving the high-use systems or wasting resources on low-utilization systems. This works by setting metadata in your host aggregates and matching extra_specs in your flavor types.

The first step is setting the aggregate metadata keys cpu_allocation_ratio and ram_allocation_ratio to a floating-point value. The filter schedulers AggregateCoreFilter and AggregateRamFilter will use those values rather than the global defaults in nova.conf when scheduling to hosts in the aggregate. It is important to be cautious when using this feature, since each host can be in multiple aggregates but should have only one allocation ratio for each resources. It is up to you to avoid putting a host in multiple aggregates that define different values for the same resource.

This is the first half of the equation. To get flavor types that are guaranteed a particular ratio, you must set the extra_specs in the flavor type to the key-value pair you want to match in the aggregate. For example, if you define extra_specs cpu_allocation_ratio to "1.0", then instances of that type will run in aggregates only where the metadata key cpu_allocation_ratio is also defined as "1.0." In practice, it is better to define an additional key-value pair in the aggregate metadata to match on rather than match directly on cpu_allocation_ratio or core_allocation_ratio. This allows better abstraction. For example, by defining a key overcommit and setting a value of "high," "medium," or "low," you could then tune the numeric allocation ratios in the aggregates without also needing to change all flavor types relating to them.

Note

Previously, all services had an availability zone. Currently, only the nova-compute service has its own availability zone. Services such as nova-scheduler, nova-network, and nova-conductor have always spanned all availability zones.

When you run any of the following operations, the services appear in their own internal availability zone (CONF.internal_service_availability_zone):

• nova host-list (os-hosts)
• euca-describe-availability-zones verbose
• nova service-list

The internal availability zone is hidden in euca-describe-availability_zones (nonverbose).

CONF.node_availability_zone has been renamed to CONF.default_availability_zone and is used only by the nova-api and nova-scheduler services.

CONF.node_availability_zone still works but is deprecated.

Scalable Hardware

While several resources already exist to help with deploying and installing OpenStack, it's very important to make sure that you have your deployment planned out ahead of time. This guide presumes that you have at least set aside a rack for the OpenStack cloud but also offers suggestions for when and what to scale.

Hardware Procurement

“The Cloud” has been described as a volatile environment where servers can be created and terminated at will. While this may be true, it does not mean that your servers must be volatile. Ensuring that your cloud's hardware is stable and configured correctly means that your cloud environment remains up and running. Basically, put effort into creating a stable hardware environment so that you can host a cloud that users may treat as unstable and volatile.

OpenStack can be deployed on any hardware supported by an OpenStack-compatible Linux distribution.

Hardware does not have to be consistent, but it should at least have the same type of CPU to support instance migration.

The typical hardware recommended for use with OpenStack is the standard value-for-money offerings that most hardware vendors stock. It should be straightforward to divide your procurement into building blocks such as "compute," "object storage," and "cloud controller," and request as many of these as you need. Alternatively, should you be unable to spend more, if you have existing servers—provided they meet your performance requirements and virtualization technology—they are quite likely to be able to support OpenStack.

Capacity Planning

OpenStack is designed to increase in size in a straightforward manner. Taking into account the considerations that we've mentioned in this chapter—particularly on the sizing of the cloud controller—it should be possible to procure additional compute or object storage nodes as needed. New nodes do not need to be the same specification, or even vendor, as existing nodes.

For compute nodes, nova-scheduler will take care of differences in sizing having to do with core count and RAM amounts; however, you should consider that the user experience changes with differing CPU speeds. When adding object storage nodes, a weight should be specified that reflects the capability of the node.

Monitoring the resource usage and user growth will enable you to know when to procure. ops_logging_monitoring details some useful metrics.

Burn-in Testing

The chances of failure for the server's hardware are high at the start and the end of its life. As a result, dealing with hardware failures while in production can be avoided by appropriate burn-in testing to attempt to trigger the early-stage failures. The general principle is to stress the hardware to its limits. Examples of burn-in tests include running a CPU or disk benchmark for several days.