Chapter 4. Architecture Examples

Short-lived workloads can include continuous integration and continuous deployment (CI-CD) jobs, which create large numbers of compute instances simultaneously to perform a set of compute-intensive tasks. The environment then copies the results or the artifacts from each instance to long-term storage before it terminates the instances.

Long-lived workloads, such as a Hadoop cluster or an HPC cluster, typically receive large data sets, perform the computational work on those data sets, and then push the results to long-term storage. When the computational work ends, the instances are idle until they receive another job. Environments for long-lived workloads are often larger and more complex, but you can offset the cost of building these environments by keeping them active between jobs.

Workloads in a compute-focused OpenStack cloud generally do not require persistent block storage, except some uses of Hadoop with HDFS. A shared file system or object store maintains initial data sets and serves as the destination for saving the computational results. By avoiding input-output (IO) overhead, you can significantly enhance workload performance. Depending on the size of the data sets, you might need to scale the object store or the shared file system.

Cloud users expect instant access to new resources as needed. Therefore, you must plan for typical usage and for sudden spikes in resource demand. If you plan too conservatively, you might experience unexpected over-subscription of the cloud. If you plan too aggressively, you might experience unexpected underutilization of the cloud unnecessary operations and maintenance costs.

The key factor in expansion planning is analytics of trends in cloud usage over time. Measure the consistency with which you deliver services instead of the average speed or capacity of the cloud. This information can help you model capacity performance and determine the current and future capacity of the cloud.

For information about monitoring software, see Section 3.9, “Additional Software”.

Typical server offerings today include CPUs with up to 12 cores. In addition, some Intel CPUs support Hyper-Threading Technology (HTT), which doubles the core capacity. Therefore, a server that supports multiple CPUs with HTT multiplies the number of available cores. HTT is an Intel proprietary simultaneous multi-threading implementation that is used to improve parallelization on the Intel CPUs. Consider enabling HTT to improve the performance of multi-threaded applications.

The decision to enable HTT on a CPU depends on the use case. For example, disabling HTT can help intense computing environments. Running performance tests of local workloads with and without HTT can help determine which option is more appropriate for a particular case.

You can add capacity to your compute environment with one or more of the following strategies:

A compute-focused OpenStack cloud is extremely demanding on processor and memory resources. Therefore, you should prioritize server hardware that can offer more CPU sockets, more CPU cores, and more RAM.

Network connectivity and storage capacity are less critical to this architecture. The hardware must provide enough network connectivity and storage capacity to meet minimum user requirements, but the storage and networking components primarily load data sets to the computational cluster and do not require consistent performance.

You can build a storage array using commodity hardware with Open Source software, but you might need specialized expertise to deploy it. You can also use a scale-out storage solution with direct-attached storage in the servers, but you must ensure that the server hardware supports the storage solution.

Consider the following factors when you design your storage hardware:

In addition to basic network considerations described in Chapter 2, Networking In-Depth, consider the following factors:

Monitoring and alerting services are critical in cloud environments with high demands on storage resources. These services provide a real-time view into the health and performance of the storage systems. An integrated management console, or other dashboards that visualize SNMP data, helps to discover and resolve issues with the storage cluster.

A storage-focused cloud design should include:

Consider using the following external network components:

Although OpenStack Networking provides a tunneling feature, it is restricted to networking-managed regions. To extend a tunnel beyond the OpenStack regions to another region or to an external system, implement the tunnel outside OpenStack or use a tunnel-management system to map the tunnel or the overlay to an external tunnel.

Some workloads require a larger MTU due to the transfer of large blocks of data. When providing network service for applications such as video streaming or storage replication, configure the OpenStack hardware nodes and the supporting network equipment for jumbo frames wherever possible. This configuration maximizes available bandwidth usage.

Configure jumbo frames across the entire path that the packets traverse. If one network component cannot handle jumbo frames, the entire path reverts to the default MTU.

If you need floating IPs instead of fixed public IPs, you must use NAT. For example, use a DHCP relay mapped to the DHCP server IP. In this case, it is easier to automate the infrastructure to apply the target IP to a new instance, instead of reconfiguring legacy or external systems for each new instance.

The NAT for floating IPs that is managed by OpenStack Networking resides in the hypervisor. However, other versions of NAT might be running elsewhere. If there is a shortage of IPv4 addresses, you can use the following methods to mitigate the shortage outside of OpenStack:

In some cases it may be desirable to use only IPv6 addresses on instances and operate either an instance or an external service to provide a NAT-based transition technology such as NAT64 and DNS64. This configuration provides a globally-routable IPv6 address, while consuming IPv4 addresses only as necessary.

QoS impacts network-intensive workloads because it provides instant service to packets with high priority because of poor network performance. In applications such as Voice over IP (VoIP), differentiated service code points are usually required for continued operation.

You can also use QoS for mixed workloads to prevent low-priority, high-bandwidth applications such as backup services, video conferencing, or file sharing, from blocking bandwidth that is needed for the continued operation of other workloads.

You can tag file-storage traffic as lower class traffic, such as best effort or scavenger, to allow higher-priority traffic to move through the network. In cases where regions in a cloud are geographically distributed, you might also use WAN optimization to reduce latency or packet loss.

The routing and switching architecture should accommodate workdloads that require network-level redundancy. The configuration depends on your selected network hardware, on the selected hardware performance, and on your networking model. Examples include Link Aggregation (LAG) and Hot Standby Router Protocol (HSRP).

The workload also impacta the effectiveness of overlay networks. If application network connections are small, short lived, or bursty, running a dynamic overlay can generate as much bandwidth as the packets the network carries. Overlay can also induce enough latency to cause issues with the hypervisor, which causes performance degradation on packet-per-second and connection-per-second rates.

By default, overlays include a secondary full-mesh option that depends on the workload. For example, most web services applications do not have major issues with a full-mesh overlay network, and some network monitoring tools or storage replication workloads have performance issues with throughput or excessive broadcast traffic.