Table 1.

Main categories of high-performance computing platforms

Large-scale computing platform	computing architectures	Advantages	Disadvantages	example applications
Cluster computing	Multiple computers linked together, typically through a fast local area network, that effectively function as a single computer	Cost-effective way to realize supercomputer performance	Requires a dedicated, specialized facility, hardware, system administrators and IT support	BLAST Bayesian network reconstruction Computing genetic associations in large-scale GWA studies
Cloud computing	Computing capability that abstracts the underlying hardware architectures (for example, servers, storage and networking), enabling convenient, on-demand network access to a shared pool of computing resources that can be readily provisioned and released (NIST Technical Report)	The virtualization technology used results in extreme flexibility; good for one-off HPC tasks, for which persistent resources are not necessary	Privacy concerns; less control over processes; bandwidth is limited as large data sets need to be moved to the cloud before processing	Searching sequence databases Aligning raw sequencing reads to genomes General purpose genomics tools (for example, GeneSifter from Geospiza) Most applications running on a cluster can be transferred to a cloud
Grid computing	A combination of loosely coupled networked computers from different administrative centres that work together on common computational tasks. Typified by volunteer computing efforts (such as Folding@Home), which ‘scavenge’ spare computational cycles from volunteers’ computers	Ability to enlist large-scale computational resources at low or no cost (large-scale volunteer-based efforts)	Big data transfers are difficult or impossible; minimal control over underlying hardware, including availability	Protein folding (Folding@Home) Proteome analysis Protein prediction (Rosetta@Home) Predicting interactions between small molecules and proteins (FightAIDS@Home) Condor project
Heterogeneous computing	Computers that integrate specialized accelerators — for example, GPUs or reconfigurable logic (FPGAs) — alongside GPPs	Cluster-scale computing for a fraction of the cost of a cluster; optimized for computationally intensive fine-grained parallelism; local control of data and processes	Significant expertise and programmer time required to implement applications; not generally available in cluster- and cloud-based services	Bayesian network learning Protein folding (Folding@Home) Molecular dynamics simulation (NAMD) BLAST CLUSTALW HMMER Reconstruction of evolutionary trees

The above categories are not exclusive. For example, heterogeneous computers are often used as the building blocks of cluster, grid or cloud computing systems; the shared computational clusters available in many organizations could be described as private Platform as a Service (PaaS) clouds. The main differences between the platforms are degree of coupling and tenancy — grid and cloud computers are designed for loosely coupled parallel workloads, with the grid resources allocated exclusively for a single user, whereas the underlying hardware resources in the cloud are typically shared among many users (multi-tenancy). Cluster computers are typically used for tightly coupled workloads and are often allocated to a single user. FPGA, field-programmable gate array; GPP, general purpose processor; GPU, graphics processing unit; GWA, genome-wide association; HPC, high-performance computing; NIST, National Institute of Standards and Technology.