Definitions - High Performance Computing Lens

Definitions

The AWS Well-Architected Framework is based on six pillars: operational excellence, security, reliability, performance efficiency, cost optimization, and sustainability. When architecting solutions, you make tradeoffs between pillars based upon your business context. These business decisions can drive your engineering priorities. You might reduce cost at the expense of reliability in development environments, or for mission-critical solutions, you might improve reliability with increased costs. Security and operational excellence are generally not traded off against other pillars.

  • HPC workloads: HPC users commonly speak of jobs. AWS refers to these as workloads.

  • Loosely coupled: Loosely coupled workloads typically see a large number of small jobs working independently of one another. This is sometimes referred to as high-throughput computing (HTC).

  • Tightly coupled: Tightly coupled workloads, such as Computational Fluid Dynamics on AWS and weather forecasting, have multi-threaded or multi-core processes working together. For additional information on these use cases, see the Tightly Coupled Scenarios.

  • Workload scheduling: Job scheduling is a means of queuing multiple jobs up to be processed (either sequentially or concurrently depending on compute resources available). Several industry-standard HPC job schedulers are available with AWS services, such as Slurm with AWS ParallelCluster, as well as AWS Partner Independent Software Vendors (ISVs) offerings. Job scheduling can also be defined using serverless algorithm design.

  • Compute vCPUs: HPC users may refer to a server as a node while AWS refers to a virtual server as an Amazon EC2 instance. AWS names and sizes instances based on the available resources allocated to the instances, such as the number of vCPUs. Depending on the instance type, a vCPU could be a thread through simultaneous multithreading (SMT) or a physical CPU core. For instance specifics, see Amazon EC2 instance type specifications.

  • Cluster placement group: A cluster placement group is a logical grouping of your compute instances within a single Availability Zone (AZ) within the same high-bisection bandwidth segment of the network. Cluster placement groups provide low latency and high bandwidth between your compute instances.

  • Elastic Fabric Adapter: Tightly coupled workloads on AWS often benefit from the use of the Elastic Fabric Adapter (EFA) and cluster placement groups. EFA is a networking technology that provides lower and more consistent latency and higher throughput than the TCP transport traditionally used in cloud-based HPC systems. EFA is described in more detail in the Networking section.

  • Amazon Machine Image (AMI): An Amazon Machine Images in Amazon EC2 is an image that contains the software for an EC2 instance to launch. These are images supported and maintained by AWS, as well as Marketplace offerings and the ability for users to create their own.

  • Blue/green instances: A blue/green deployment is a strategy whereby two separate, but identical, deployments are created. The blue environment runs the current application version while the green environment runs the new application version.

  • CI/CD: Continuous Integration and Continuous Development. CI is a set of SW coding practices that drive development teams to frequently implement small code changes and check them in to a version control repository. It establishes an automated way to build, package, and test their applications. This encourages developers to commit code changes more frequently, leading to better collaboration and code quality. CD automates application delivery to production, development, and testing environments and is an automated way to push code changes to these environments.

  • Computational Fluid Dynamics (CFD): This is a form of Computational Aided Engineering (CAE) to simulate fluid flows around or within objects. It uses a Finite Volume approach to solve numerical equations across domains split into many smaller cells in an iterative approach. A commonly used example is aerodynamics of cars.

  • Data retention requirement: Data management – where and how much data is retained – is a significant consideration for HPC environments. User and regulatory requirements for data retention should be considered, with different strategies possible.

  • EasyBuild: Third party package management software that provides a software build and install framework to manage (chiefly scientific) software on High Performance Computing (HPC) systems efficiently.

  • Fail fast: This is a philosophy for frequent, incremental, development and testing to determine whether a development path is the best one.

  • Finite Element Analysis (FEA): A form of Computational Aided Engineering (CAE) that simulates structural analysis. This is used for many engineered products such as crash simulation in cars or building reaction to earthquakes. This simulation type is heavily used in Mechanical Engineering.

  • Genomics: A branch of study that refers to mapping the genetic code of living tissues to identify the genetic material and the result of interactions of genes with each other and their environment.

  • Hypervisor: Software that allows for the running of multiple virtual machines on a single physical machine. AWS use the AWS Nitro System.

  • IOPS: Input/output operations per second. This is a measure of data throughput on a bus, in a CPU or GPU, or when using a storage device.

  • Job scheduling software: This is software that knows what resources are available and allows users to submit work to run on the compute machines. If all resources are busy, jobs are queued up until the required resources are available.

  • Lazy load: When data in an S3 bucket linked to a Data Repository Association (for FSx for Lustre) is accessed, Amazon File Cache automatically loads the metadata and file contents if they are not already present in the cache. Lazy load allows reading or writing data or metadata to files in a DRA directory.

  • Luigi: A lightweight, open-sourced, Python-based workflow scheduler that is typically used for orchestrating data assets in the Hadoop ecosystem.

  • Memory to Core Ratio: A compute machine has a physical amount of Random Access Memory (RAM) available as well as number of cores. By dividing RAM by core count it is possible to determine the amount of RAM per core available. If the machine has a job running on it that divides the compute resources equally, the workload that is going to run will be limited on a per-core basis by the amount of memory available for that core. This can place an upper limit on workload sizes on a particular machine configuration.

  • Monte Carlo simulations: The Monte Carlo simulation method is a means of modelling approximate solutions using repeated random sampling.

  • Message Passing Interface (MPI):Message Passing Interface is a standard that defines communication between machines that are working on workloads together, allowing those workloads to scale out to more than one instance while solving the same problem.

  • Network access control lists (NACLs): A Control subnet traffic with network access control lists allows administrators to control who has access into a network. This controls traffic at the subnet level rather than at an instance.

  • Parallel file system: This is a file system that allows multiple machines to access the same file system. It is also possible for multiple reads/writes to be made synchronously by a single instance to speed up file input/output.

  • Personally identifiable information (PII): Information that identifies one or more details about a person including, but not limited to, name, address, gender, birth-date, SSN, and Government identifier and has strict compliance restrictions on organizations that use it.

  • Placement group: In a physical, on-premises, High Performance Compute (HPC) cluster, machines would generally be physically in very close proximity to each other. This reduces network latency as the transmission distance between machines is low while also needing few numbers of 'hops' across different hardware (e.g. network switches). The equivalent on AWS is requesting instances in a Placement Group whereby requested instances will be allocated as close to each other as possible.

  • Recovery Point Objective (RPO): According to the US National Institute of Standards and Technology, Recovery Point Objective is defined as The point in time to which data must be recovered after an outage.

  • Recovery Time Objective (RTO): According to the US National Institute of Standards and Technology, Recovery Time Objective is defined as The overall length of time an information system's components can be in the recovery phase before negatively impacting the organization's mission or mission/business processes.

  • REST-based communication: REST is defined by the US NIST as a software architectural style that defines a common method for defining APIs for Web services.

  • Role-based access control (RBAC): the US NIST provides a definition as 'Access control based on user roles (that is a collection of access authorizations that a user receives based on an explicit or implicit assumption of a given role). Role permissions may be inherited through a role hierarchy and typically reflect the permissions needed to perform defined functions within an organization. A given role may apply to a single individual or to several individuals.'

  • SBGrid: Third party package management software that provides a software build and install framework to manage software on High Performance Computing (HPC) systems efficiently.

  • Slurm: This is an open-source workload manager (queuer) to schedule jobs for Linux or Unix based systems. Now also known as the Slurm Workload Manager, previously Simple Linux Utility for Resource Management (Slurm). This is the default choice for scheduler for AWS ParallelCluster.

  • Slurm partition: Computational resources on a cluster with a queue hosted by a Slurm system can be split across different partitions, allowing users to choose what hardware configuration to queue their jobs on to.

  • Spack: Third party package management software that provides a software build and install framework to manage (chiefly scientific) software on High Performance Computing (HPC) systems efficiently.