Cluster computing offers many benefits: high availability through fault tolerance and resiliency, load balancing and scaling capabilities, and performance improvements.
There are some important terms to keep in mind when discussing the robustness of a system:
Availability- The accessibility of a system or service over a period of time, typically expressed as a percentage of uptime in a given year (e.g., 99.).999% availability, or five nines).
Elasticity- The ability of the system to recover from failures.
Fault tolerance – the ability of a system to continue to provide services in the event of a failure.
Reliability- the probability that the system will work as expected.
Redundancy- Duplicate critical resources to improve system reliability.
Applications running on a single machine have a single point of failure, resulting in poor system reliability. If the computer hosting the application fails, there is almost always downtime during infrastructure recovery. Maintaining a certain level of redundancy helps improve reliability and can reduce the amount of time your application is unavailable. This can be achieved by preemptively running the application on a second system (which may or may not be receiving traffic) or by using the application to pre-configure a cold system if it is not currently running. These configurations are referred to as active-active and active-passive configurations, respectively. When a failure is detected, the active-active system can immediately fail over to the second machine, while the active-passive system will fail over once the second machine comes online.
A computer cluster consists of multiple nodes running the same process at the same time, so it is an active-active system. Active-active systems are typically fault-tolerant because systems are inherently designed to handle node loss. If one node fails, the remaining nodes are ready to take on the workload of the failed node. That being said, a cluster that requires a leader node should run at least two leader nodes in an active-active configuration. This prevents the cluster from becoming unavailable if the leader node fails.
In addition to being more fault-tolerant, clusters can improve resiliency by making it easy for recovered nodes to rejoin the system and bring the cluster back to optimal size. Any amount of system downtime is costly for an organization and can result in a poor customer experience, so it's critical that systems are resilient and fault-tolerant in the event of a failure. The use of clusters improves the resiliency and fault tolerance of the system, resulting in higher availability. Clusters with these characteristics are called "highly available" or "failover" clusters.
Load balancing is the act of distributing traffic among the nodes of a cluster to optimize performance and prevent any single node from receiving a disproportionate amount of work. The load balancer can be installed on the leader node or configured separately from the cluster. By performing periodic health checks on each node in the cluster, the load balancer is able to detect if a node is down, and if it does, it routes incoming traffic to other nodes in the cluster.
Although a computer cluster does not load balance itself, it can perform load balancing across its nodes. This configuration is called a "load balancing" cluster, and is usually a high-availability cluster at the same time.
There are two classifications of scaling: vertical scaling and horizontal scaling. Vertical scaling (also known as scaling up or down) involves increasing or decreasing the amount of resources allocated to a process, such as the amount of memory, the number of processor cores, or the available storage space. On the other hand, horizontal scaling (scale-out scale-out) refers to the running of additional parallel jobs on the system.
When maintaining a cluster, it's important to monitor resource usage and scale to ensure that cluster resources are being utilized appropriately. Fortunately, the nature of clustering makes horizontal scaling trivial – administrators only need to add or remove nodes as needed, and keep in mind the lowest level of redundancy to ensure that the cluster remains highly available.
In terms of parallelization, clusters can achieve higher levels of performance than single machines. This is because they are not limited by a certain number of processor cores or other hardware. In addition, horizontal scaling can prevent the system from running out of resources, thereby maximizing performance.
High-performance computing (HPC) clusters take advantage of the parallelism of computer clusters to achieve the highest possible level of performance. Supercomputers are a common example of an HPC cluster.