Throughput (computing)

Throughput (computing)
Name	Throughput (computing)
Field	Computing, Computer science
Units	bits per second, transactions per second, operations per second

Throughput (computing) is a measure of the rate at which a computing system completes useful work over time. It quantifies delivered capacity in units such as bits per second, transactions per second, or operations per second, and is central to performance analysis in systems designed by organizations like Intel Corporation, IBM, and NVIDIA Corporation. Throughput interacts with latency, availability, and reliability in systems used by institutions such as NASA, European Space Agency, and MIT research labs.

Definition and scope

Throughput denotes the sustained rate of successful work completion for a component or system and appears in contexts including networking, storage, databases, and parallel computing; it is a practical complement to theoretical peak rates advertised by vendors such as AMD and ARM Limited. Historically framed by standards bodies like IEEE and IETF, throughput is distinct from terms used in benchmarks administered by entities such as SPEC and Transaction Processing Performance Council. Practical scope spans from microarchitectural instruction throughput analyzed by researchers at Bell Labs and Carnegie Mellon University to cloud-scale throughput managed by providers like Amazon Web Services and Google LLC.

Metrics and measurement

Metrics for throughput include bits per second (bps), packets per second (pps), transactions per second (TPS), queries per second (QPS), and input/output operations per second (IOPS); these are the same classes used in reports by Microsoft and Oracle Corporation. Measurement methodologies incorporate averaging, peak sampling, and moving-window techniques defined in publications from ACM and IEEE Computer Society. Tools and suites commonly used to measure throughput include benchmarks from TPC (for database workloads), SPEC CPU and SPECjAppServer (for application servers), networking tools used by researchers at MIT Lincoln Laboratory, and measurement frameworks employed by CERN for high-energy physics data flows. Careful measurement accounts for warm-up effects, cold caches analyzed in studies at Stanford University and University of California, Berkeley, and workload characterization techniques promoted by DARPA programs.

Factors affecting throughput

Throughput depends on hardware resources such as CPU cores designed by Arm Holdings and Intel Corporation, GPU parallelism from NVIDIA Corporation, memory hierarchy influenced by designs from Samsung Electronics and SK Hynix, storage media like SSDs manufactured by Western Digital Corporation, and network interfaces created by Cisco Systems. Software factors include scheduler policies in Linux and FreeBSD, database engines such as MySQL and PostgreSQL, virtualization stacks like VMware and KVM, and container orchestration by Kubernetes and Docker, Inc.. System architecture choices—symmetric multiprocessing employed by IBM mainframes, distributed designs used by Apache Hadoop and Apache Spark, and microservices popularized by firms like Netflix—all modulate throughput. External constraints like power budgets in designs influenced by ARM partners and regulatory requirements in markets served by European Commission also shape achievable throughput.

Throughput in different computing domains

Network throughput is central in projects by Cisco Systems and standards from IETF and is measured in contexts such as backbone links studied by Verizon and AT&T. Storage throughput matters in systems built by NetApp and evaluated in research at Lawrence Berkeley National Laboratory. Database throughput is benchmarked by TPC and optimized by vendors such as Oracle Corporation and Microsoft SQL Server. High-performance computing throughput is characterized in TOP500 lists maintained by researchers at Lawrence Livermore National Laboratory and involves architectures from Cray Inc. and Hewlett Packard Enterprise. Cloud throughput considerations appear in offerings by Amazon Web Services, Google Cloud Platform, and Microsoft Azure and in academic cloud studies at University of Illinois Urbana-Champaign and Princeton University.

Optimization techniques and trade-offs

Optimizations include parallelization strategies championed by Amdahl's law proponents and researchers at Intel; pipelining and superscalar techniques pioneered in designs at Bell Labs and DEC; caching hierarchies improved by projects at IBM Research; I/O batching seen in systems by Facebook and Twitter; and protocol tuning studied by IETF working groups. Trade-offs often involve latency versus throughput debated in literature from ACM SIGOPS and IEEE Transactions on Computers, consistency versus throughput in the CAP theorem discussed by academics at University of California, Berkeley, and energy efficiency versus throughput addressed in initiatives by Green Grid and Energy Star. Load balancing approaches from HAProxy and Nginx and scheduling algorithms developed at Carnegie Mellon University illustrate operational trade-offs in production systems used by firms like Spotify and Airbnb.

Historical development and notable benchmarks

The concept of throughput evolved alongside computing milestones such as the development of the ENIAC and the rise of mainframes from IBM; advances continued through microprocessor scaling by Intel and the emergence of networking standards like Ethernet from Xerox PARC. Notable benchmarks include SPECint and SPECfp from SPEC, TPC-C and TPC-H from Transaction Processing Performance Council, LINPACK used in the TOP500 list, and database and web benchmarks adopted by companies like Yahoo! and Google LLC. Academic milestones in throughput measurement are documented in papers from ACM SIGCOMM, USENIX, and IEEE INFOCOM, while governmental and research facilities such as Los Alamos National Laboratory and CERN have driven high-throughput system designs.

Category:Computer performance