KWEST: Kubernetes Workload Simulator and Re- play Tool
Date Issued
May 2026
Author(s)
Paraskeva, Matthew
Advisor
Abstract
Evaluating Kubernetes scheduling behavior for different workloads is costly and difficult to reproduce on
real clusters due to infrastructure overhead, environmental noise, and inconsistent initial conditions. Al-
though simulation provides controlled execution, simulator-based workflows commonly rely on manual
file handling, incomplete provenance capture, and fragmented result inspection, which limits repeatabil-
ity and auditability.
This thesis extends the KWEST simulator into a practical experimentation platform for Kubernetes-
oriented workload studies. The platform introduces a workload pipeline that supports both synthetic
scenario generation and trace-driven replay derived from Kubernetes pod lifecycle events, together with
a cluster import mechanism that exports and ingests Kubernetes node snapshots reflecting scheduler-
visible resources (e.g. allocatable CPU and memory). A web-based interface integrates configuration
creation and validation, run execution, and run history navigation. To support repeatability and controlled
comparisons, the platform persists run metadata, referenced inputs, and produced artifacts in Couchbase
under a stable runId, using gzip+base64 encoding to reduce storage overhead, while report views provide
structured access to key simulator outputs.
A scenario-based evaluation assesses trace correctness, artifact completeness, reproducibility under fixed
inputs, controlled cross-scheduler comparisons, and runtime/storage overhead. Results show that Kubernetes-
derived inputs can be captured and reused consistently, that run provenance is sufficient to repeat exper-
iments, and that orchestration overhead remains modest relative to simulator runtime, with compression
substantially reducing stored artifact size
real clusters due to infrastructure overhead, environmental noise, and inconsistent initial conditions. Al-
though simulation provides controlled execution, simulator-based workflows commonly rely on manual
file handling, incomplete provenance capture, and fragmented result inspection, which limits repeatabil-
ity and auditability.
This thesis extends the KWEST simulator into a practical experimentation platform for Kubernetes-
oriented workload studies. The platform introduces a workload pipeline that supports both synthetic
scenario generation and trace-driven replay derived from Kubernetes pod lifecycle events, together with
a cluster import mechanism that exports and ingests Kubernetes node snapshots reflecting scheduler-
visible resources (e.g. allocatable CPU and memory). A web-based interface integrates configuration
creation and validation, run execution, and run history navigation. To support repeatability and controlled
comparisons, the platform persists run metadata, referenced inputs, and produced artifacts in Couchbase
under a stable runId, using gzip+base64 encoding to reduce storage overhead, while report views provide
structured access to key simulator outputs.
A scenario-based evaluation assesses trace correctness, artifact completeness, reproducibility under fixed
inputs, controlled cross-scheduler comparisons, and runtime/storage overhead. Results show that Kubernetes-
derived inputs can be captured and reused consistently, that run provenance is sufficient to repeat exper-
iments, and that orchestration overhead remains modest relative to simulator runtime, with compression
substantially reducing stored artifact size
Affiliation
File(s)![Thumbnail Image]()
Name
Matthew Paraskeva.msc.2026.abstract.pdf
Size
142.04 KB
Format
Adobe PDF
Checksum (MD5)
54862f739dc7445fd06478d46eaf0eee