Types of Workloads
2026-03-08 22:41
Status:
Tags: #software #statistics #performance
Types of Workloads
The level of detail in workload evaluation strongly depends on the goals of said evaluation; within an operating system scheduler frame of reference, it is sufficient to consider a process as "computing" but we will need more details to evaluate the CPU architecture and instruction set.
Dynamic vs. Static Workloads
A key property of workloads is their rate of events — how many operations occur per unit time.
- A desktop system may process hundreds of user commands per day.
- A supercomputer may execute only hundreds of large jobs per day.
- A network router may process billions of packets per day.
The important distinction is between size and rate:
- Size describes a fixed amount of work.
- Rate describes work arriving continuously over time.
This distinction leads to the concepts of static and dynamic workloads.
Static workloads
A static workload consists of a fixed set of jobs known in advance.
The system simply processes the given work until completion.
Example:
- Running several copies of applications
, , and - Benchmark suites where the workload is predetermined
Static workloads are easier to evaluate, but often less realistic.
Dynamic workloads
A dynamic workload represents continuous arrival of work over time.
The system is never finished; it is always handling incoming tasks.
Example:
- user commands on a desktop
- incoming packets on a router
- job submissions on a cluster
Implications for performance evaluation
- Workload mix must evolve over time
Evaluating dynamic systems requires modeling:
- possible job types
- their popularity
- how they appear over time
Static benchmarks avoid this complexity.
- Arrival processes must be modeled
Dynamic workloads introduce an additional stochastic component:
Examples include Poisson arrivals or bursty traffic patterns.
Static workloads assume all jobs are already present.
- Observed distributions can differ from input distributions
Sampling the running system can give a misleading picture.
Example:
- Input distribution: many short jobs, few long jobs.
- System snapshot: long jobs appear more common.
This occurs because long jobs remain in the system longer, increasing the probability of observing them.
- System state matters (aging effects)
Performance depends on the system’s current state.
A system that has processed many jobs may experience:
- memory fragmentation
- cache pollution
- resource contention
Static workloads typically start with a clean system, hiding such effects.
- Some phenomena require dynamic workloads
Certain behaviors only emerge through long-term interaction of many jobs:
- memory leaks
- software aging
- paging thrashing
- long-term resource fragmentation
- Static workloads cannot reveal these effects.
TLDR
Static workloads simplify experimentation but may miss essential system behaviors, whereas dynamic workloads better capture real system operation, including arrival processes, system aging, and long-term interactions between tasks.