Workload representativeness and trace provenance

Start with the decision population: tenants, request classes, time periods, regions, data sizes, and arrival conditions. Record collection location, clock source, missing intervals, sample probabilities when known, and the software schema that assigned meaning to each field. Preserve immutable raw-trace identifiers and a transformation manifest for filtering, anonymization, deduplication, and unit conversion. A privacy-preserving transformation may intentionally remove correlations that matter to cache locality or routing. Document that loss instead of calling the transformed trace behaviorally identical.

Compare collected composition with the target rather than assuming a large file is representative. Oversampling rare slow operations can improve diagnostic coverage but distort an unweighted mean. A workload with many cheap lookups and occasional expensive scans needs stated population weights if the claim concerns mean request cost. Compute weighted costs in their original units before forming an overall ratio. The resulting aggregate does not replace per-class results, because a favorable mean can conceal a regression in a tenant or a latency-sensitive operation.

An open-loop replay schedules arrivals from the trace independently of response completion. A closed-loop client submits subsequent work after earlier responses, possibly with think time. These designs can expose very different overload behavior. Dropping arrivals that cannot be issued on time may hide queueing, while sending them all in a burst creates a different workload. Preserve intended and actual issue times, completion times, errors, and admission decisions. Include the load generator in capacity checks so a client bottleneck is not mistaken for server saturation.

Time compression changes more than the label on an axis. Uniformly dividing interarrival gaps by a factor c multiplies the intended average arrival rate by c for the same event count and proportionally shortened observation interval. It need not preserve cache residency, background maintenance overlap, or user adaptation. Synthetic scaling should also state whether keys, object sizes, tenant counts, and temporal correlations change. A response-dependent production trace cannot generally be replayed under a different system while retaining all feedback relationships.

List information that is unavailable: requests rejected before instrumentation, cache state before collection, traffic from excluded tenants, or deleted identifiers. Separate a convenience sample from a probabilistic sample; unknown selection probabilities can prevent justified population reweighting. Use multiple collection windows and sensitivity analyses to show how conclusions depend on composition. When a representative proprietary trace cannot be published, release a permitted surrogate and explain which properties it matches and which conclusions an independent investigator cannot reproduce with it.

The cache instrument shows why stream structure matters even with the same request count. Reset, set both C=4 and s=4, predict four misses and thirteen hits after one Step, then Step once. Separately Reset, set both C=4 and s=1, predict seventeen misses after one Step, then Step once. Both runs start empty and evaluate seventeen requests to address (i·s) mod 16. This fully associative LRU toy isolates a synthetic locality difference; it is not a production trace, processor cache benchmark, or validation of a real replay generator.