01Introduction to Cloud Computing
The U.S. National Institute of Standards and Technology defines cloud computing as a model for enabling convenient, on-demand network access to a shared pool of configurable computing resources that can be rapidly provisioned and released with minimal management effort.
The NIST definition (Special Publication 800-145, 2011) identifies five essential characteristics: on-demand self-service, broad network access, resource pooling, rapid elasticity and measured service. It also names four deployment models: public, private, community and hybrid cloud.
02Distributed Computing Fundamentals
A distributed system is a collection of independent computers that communicate over a network and appear to users as a single coherent system. Distribution is used to gain capacity, to keep working when individual machines fail, and to place computation close to users or data.
Distributed systems are hard because of three facts: messages can be delayed or lost, machines can fail independently, and there is no single global clock. Peter Deutsch's well-known “fallacies of distributed computing” list the assumptions newcomers tend to make wrongly — that the network is reliable, latency is zero and bandwidth is infinite, among others.
03Cloud Service Models
| Layer | IaaS | PaaS | SaaS |
|---|---|---|---|
| Application and data | Customer | Customer | Provider (customer owns data) |
| Runtime and middleware | Customer | Provider | Provider |
| Operating system | Customer | Provider | Provider |
| Virtualization, servers, storage, network | Provider | Provider | Provider |
The further up the stack, the less the customer manages and the less control it has. This division is known as the shared responsibility model, and it matters especially for security.
04Virtualization
Virtualization lets one physical server run several isolated virtual machines (VMs). A hypervisor allocates CPU, memory and devices to each VM. Type 1 hypervisors run directly on hardware; type 2 hypervisors run on top of a host operating system. Each VM has its own full operating system, which gives strong isolation at the cost of memory and start-up time.
05Containers and Orchestration
Containers package an application with its libraries and share the host operating system's kernel, isolated using kernel features such as namespaces and control groups on Linux. They start in seconds and use fewer resources than VMs, but isolation between containers is weaker than between VMs.
Orchestration systems such as Kubernetes schedule containers across a cluster, restart failed ones, scale replicas up and down, and route network traffic. In research computing, containers are valuable for reproducibility: an analysis environment can be captured and rerun exactly.
| Virtual machines | Containers | |
|---|---|---|
| Includes | Full guest operating system | Application and libraries only |
| Start-up | Typically tens of seconds or more | Typically seconds or less |
| Isolation | Strong (hardware-level) | Weaker (shared kernel) |
| Typical use | Mixed operating systems, strong tenant separation | Microservices, reproducible environments |
06Distributed Storage
Distributed storage spreads data across many machines for capacity and resilience. Block storage provides virtual disks; file storage offers shared hierarchical file systems, including parallel file systems used in HPC; object storage stores immutable objects addressed by key and scales to very large datasets. Data are usually replicated (several full copies) or protected by erasure coding (data plus parity fragments), trading storage overhead against recovery cost.
07Scalability and Elasticity
Scalability is the ability to handle more load by adding resources. Vertical scaling uses bigger machines; horizontal scaling adds more machines, which is more flexible but requires the application to work in parallel. Elasticity means resources are added and removed automatically as demand changes, so capacity follows load. Stateless services scale horizontally most easily; stateful components such as databases need partitioning (sharding) and replication.
08Fault Tolerance
In a large system, something is always failing. Fault-tolerant design assumes this and continues to provide service. Techniques include redundancy across machines and data centres, health checks with automatic failover, timeouts and retries with exponential back-off, checkpointing long computations so they can resume, and consensus protocols such as Paxos and Raft, which let a group of servers agree on a sequence of decisions as long as a majority are working.
09Parallel Processing
Parallel processing divides work so it can be done simultaneously. Data parallelism applies the same operation to different parts of a dataset; task parallelism runs different operations at once. Shared-memory programming (for example, OpenMP) suits cores on one machine; message passing (for example, MPI) suits clusters. Frameworks such as MapReduce and Apache Spark apply data parallelism to large-scale analytics. Speed-up is limited by the serial part of the work and by communication overhead.
10Data Consistency
When data are replicated, replicas can disagree. Strong consistency (linearizability) makes the system behave as though there were a single copy. Eventual consistency allows temporary differences that resolve once updates stop.
Brewer's conjecture, proved by Gilbert and Lynch in 2002, states that when a network partition occurs, a distributed system must choose between consistency and availability. In practice the choice is not all-or-nothing; many systems let designers tune consistency per operation, and latency is an equally important trade-off when there is no partition.
11Cloud Security
Cloud security follows the shared responsibility model: the provider secures the underlying infrastructure, while the customer secures its configurations, identities, data and applications. Many cloud incidents result from customer-side misconfiguration, such as storage made publicly readable by mistake. Core controls include least-privilege identity and access management, encryption in transit and at rest, network segmentation, logging and continuous configuration review. These are explored further in cybersecurity research.
12Scientific Computing Workloads
Scientific workloads vary. Tightly coupled simulations exchange data constantly and need low-latency interconnects, traditionally found in dedicated HPC clusters. Loosely coupled or “embarrassingly parallel” workloads — parameter sweeps, Monte Carlo ensembles, genome analyses — split into independent tasks and suit cloud resources well. Workflow systems chain these tasks and record provenance so results can be reproduced. Data movement is often the hidden cost: transferring large datasets can take longer than computing on them.
13Advantages and Trade-Offs
| Advantage | Corresponding trade-off |
|---|---|
| Capacity on demand without buying hardware | Costs can grow unexpectedly without monitoring |
| Resilience through redundancy | Greater design and operational complexity |
| Global reach and managed services | Dependence on a provider; migration can be difficult |
| Horizontal scalability | Consistency and coordination become harder |
| Reproducible environments with containers | Shared infrastructure requires careful security configuration |
Further reading
- Mell, P., & Grance, T. (2011). The NIST Definition of Cloud Computing (NIST Special Publication 800-145). National Institute of Standards and Technology.
- Gilbert, S., & Lynch, N. (2002). Brewer's conjecture and the feasibility of consistent, available, partition-tolerant web services. ACM SIGACT News, 33(2), 51–59.
- Ongaro, D., & Ousterhout, J. (2014). In search of an understandable consensus algorithm. Proceedings of the 2014 USENIX Annual Technical Conference.
- Kleppmann, M. (2017). Designing Data-Intensive Applications. O'Reilly Media.