high performance computing. Implies a higher percentage of CPU and memory usage than typical administrative computing, or implies a program too large for, or that takes too long on, a desktop computer. In academia, used for and implies computing for research. Also HTC, high throughput computing, essentially similar but oriented to processing large data files.
a single computer in a box, functionally similar to a desktop computer but typically more powerful and packaged for rackmount in a datacenter. Usually two CPU sockets or four sockets with very large memory vs. one socket for a desktop. Implies shared memory for the programs running on the node.
a group of nodes connected by a network. Depending on budget and need for communication, the network may be inexpensive commodity (Gigabit ethernet) or faster and more expensive (Infiniband or 40Gb Ethernet) or in between (10Gb Ethernet).
a large and powerful cluster. Exact definition uncertain.
a cluster node dedicated to scientific computing tasks. Usually controlled by a scheduler program. Usually isolated from the public internet by head nodes. Usually has a high-speed network for running parallel computing and delivering files over the network.
a cluster node connected to the public internet and dedicated to logins, editing, moving data, submitting jobs.
a cluster node dedicated to delivering networked file systems to a cluster.
a cluster node connected to the public internet and dedicated to moving data to/from external computers.
Circa 1980, one computer= one node= one CPU= one socket = one core = one processor = usually one thread. Since then computer design has become more complicated. Typical server computers used for HPC have two to eight sockets (usually two) on a mainboard or motherboard. Each socket holds a CPU “central processing unit” chip. Each CPU has 1 to 64 cores (or processors), each of which can run an independent program. Intel computers also have 1 or 2 hardware threads or hyperthreads per core, each capable of running an independent program, but which do not add any more hardware capability to the system.
a graphical processing unit, a specialized type of CPU derived from a graphics card. Effectively has hundreds of small cores. For certain tasks (those that can be effectively parallelized), is much faster than a general-purpose CPU. Presently available versions are “co-processors” that must be attached to a PCI bus which is connected to and controlled by a CPU. Also Xeon Phi, a CPU with many small cores which has the software interface of a general purpose CPU, but is used on highly parallel codes like a GPU. Xeon Phi is available either as co-processor or as a CPU.
In software, a program that runs multiple tasks or software threads, each of which sees the same available memory available from the operating system, and shares that memory using one of the multiple shared memory/multi-threading communication methods (OpenMP, pthreads, POSIX shm, MPI over shared memory, etc.). Shared memory programs cannot run across multiple nodes. In hardware, usually a node or a single computer system that supports shared memory access across its memory. Implies a limit (a little less than the amount of memory in the node) to the memory size of the running program.
In software, a program or group of programs that run on multiple nodes or shared-memory instances and use programs such as MPI to communicate between the nodes. In hardware, a cluster that runs distributed-memory programs. Distributed-memory programs are limited in memory size only by the size of the cluster that runs them, though there may be some useful limit that is smaller.
message passing interface, software standard used for most programs that use distributed memory. MPI calls lower-level functions, either networking or shared memory, so it may be used transparently over either distributed or shared memory without changing the user interface. On a cluster that means it can run transparently either on one node or multiple nodes. MPI has multiple implementations (OpenMPI, MVAPICH, several commercial) that must be used consistently to both compile and run an MPI program. Each MPI task is a separate program which can be single-threaded or multi-threaded.
A software program that cannot take advantage of multi-threading because it was written without multi-threading support. Essentially can use only one core on one node regardless of the number of cores available. Multiple single-threaded programs can be run on a single node on multiple cores if sufficient memory is available.
A design element used to make fast computers affordable. Memory is arranged in levels with very small and very fast and very expensive levels close to the CPU, and each succeeding level is made larger and slower. Most modern computers have registers (very fast and of KB size), L1 to L3 or L4 cache of MB size, and main memory of GB size, or “memory” if unspecified. The operating system automatically handles staging data from main memory through the cache and registers, unless the programmer uses assembly language to control that staging. This process makes sequential access to main memory relatively fast, as large blocks of memory can be staged through the cache while computing is ongoing, but random access to main memory is relatively slow, as the processor can idle for 200 cycles while waiting for a single element of main memory.
By analogy with memory hierarchy, the practice of using multiple disk storage systems with an HPC system. Each tier of storage is larger and slower than the preceding tier. The first “scratch” tier is relatively small and fast for a disk, usually composed of SSD, and does most direct data movement to the compute nodes. The last tier may be tape or large and inexpensive disk drives and holds longer term and larger files.
A temporary file system, designed for speed rather than reliability, and the first tier in the storage hierarchy. Usually composed of faster disks, currently SSD.
Solid state disk, memory chips packaged with an interface that appears to the computer to be a disk drive. Faster than rotating disk drives and still more expensive, though decreasing in price over time.
delay, or the time it takes to access a minimal message over a given network. Used to characterize networks in combination with
the amount of data that can be moved over a network per second.
a program running on a node that emulates a computer and connects the host computer's resources to the emulated computer. The VM runs an operating system, which runs user programs, like a physical computer. Useful for programs that do not consume a lot of CPU time, and also useful to keep user programs from exceeding memory limits, and providing a way to save the state of a user program. A single powerful computer can run a number of VMs. Also container as in Docker, a form of VM that is less isolated from the host computer than is a full VM.
A program that maintains a list of batch jobs to be executed on a cluster, ranks them in some priority order, and executes batch jobs on compute nodes as they become available. It tries to keep the cluster from being overloaded or idle. Also OS scheduler, a part of the kernel that runs on a shared-memory node and allows competing user programs access to CPU time, and IO scheduler, another part of the kernel that orders multiple disk accesses for a node.
A program that is either multi-task (like MPI) or multi-threaded (like OpenMP) or both, in order to effectively use more cores and more nodes and get more computing done. May be either shared-memory or distributed-memory. Opposite, a serial program.
The efficiency of a parallel program, usually defined as the parallel speedup of the program divided by the number of cores occupied. Speedup is defined as the serial run time divided by the parallel run time. Usually parallel computing introduces overhead, and scaling is less than 1 or 100%. Rarely, running on multiple CPUs can make each task fit within the memory cache of each CPU, avoiding waiting for main memory access, and scaling can exceed 1. In most cases, scaling starts at 1 at 1 core (by definition) and decreases as more cores are added, until some point is reached at which adding more cores adds overhead and makes the program slower.