An impact-aware policy in Slasher kept expected total cost near zero at a 15% power-reduction goal in a discrete-event simulation. The same comparison reported fewer high-priority services affected than the one-pass variant.
The 15% figure is a power-reduction goal in the simulation, and the near-zero figure is the model’s expected total cost. Slasher is presented as a general system for modulating datacenter power while meeting power targets and minimizing negative impact on hosted workloads.
The evaluated shutdown algorithm orders servers by increasing aggregate impact per Watt. It then recalculates the remaining-server scores after selected servers affect overlapping services, allowing the ranking to change during shutdown.
The capacity trade-off
One part of the work examined CPU-frequency controls. The sensitivity study covered ten request-serving applications running in CPU virtual machines without attached accelerators. Across those applications, median retained serving capacity, the share of peak capacity still available, was 89% at 10% lower dynamic power, 79% at 20%, 64% at 30% and 48% at 40%. At the deepest target, results ranged from 7% to 57%.
Vault illustrates how uneven the trade-off can be. At 2.5 GHz, the workload supported 600 requests per second, or 46% of its peak load of 1,300 requests per second. The figure sits within the ten-application CPU-only sample.
What deployment changes might provide
Another route in the paper was to block new deployments and examine the projected change in unallocated servers and occupancy. Lever characterization used Microsoft production data from 20 datacenters over nine days, including CPU-compute server power consumption. In a projection based on 140 datacenter-day examples from seven days across the same 20 datacenters, blocking new deployments corresponded within two hours to a median 2% increase in unallocated servers and roughly 8% lower occupancy. Past 12 hours, the text reports 3% to 7% additional unallocated servers and a 5% to 15% occupancy reduction.
In the shutdown simulation, the first 5% of power reduction was attributed to shutting down excess unallocated servers without violating buffer constraints. After that initial step, the algorithm ranked remaining servers by aggregate impact per Watt and recalculated the scores whenever a selected server affected overlapping services.
How the policy was scored
Slasher’s workload-impact model uses conditional value at risk, also called expected shortfall, to estimate expected un-served load when service capacity falls short. In everyday terms, it gives the simulator a common way to compare shutdown choices by estimating the load left unserved in the model.
Stratosim replays power and resource telemetry through modeled entities ranging from the grid connection to virtual machines and represents communication delays. For the shutdown-policy comparison, the input was a production datacenter trace with 30 days of aggregate service-load history, one hour of VM-attributed power telemetry and one hour of VM-allocation records.
At the 15% goal, AscScore-ReEval kept expected total cost near zero, while AscScore-OnePass produced non-zero impacts beyond service slack. The score variants mainly selected spot-VM or slack-capacity servers and affected fewer high-priority services.
The reported timing came from a sampled datacenter with about 5,000 non-empty servers and 10,000 services. Initial score evaluation took 4.5 to 6 seconds, and re-evaluation added 0.1 to 2 seconds. Those figures describe that sampled case.
A pilot alongside the modeling
The paper also describes a live-datacenter pilot combining a hardware controller, a software-configurable UPS, or uninterruptible power supply, and an IT controller. The setup was used to mimic a prolonged fast-frequency-response event requiring a sub-second response sustained for several minutes.
Development was ongoing: the core Slasher mechanism was operational and in production, while some levers and scenarios were still being integrated. The paper states that Stratosim would be open-source with publication.
The manuscript is an arXiv preprint, version 1, dated 26 Aug 2026. Its 15% result belongs to the Stratosim policy comparison; the pilot description concerns the control architecture and response scenario.
The quantitative results come from distinct sources: Microsoft production traces, ten CPU-only applications and a production trace used for shutdown-policy evaluation, alongside the live-pilot description. Slasher is presented as a general system, but the reported figures are tied to those traces, tested applications and modeled scenarios.
Paper data and sources
Original title: Slasher: Power Flexibility for Cloud Datacenters
Authors: Liuzixuan Lin, Fiodar Kazhamiaka, Alok Gautam Kumbhare et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text