Source-linked AI summary
FissionReady: Joint Workload and Power Scheduling for Data Centers Powered by Small Modular Reactors
Raghavendra Kanakagiri, Rohan Basu Roy, Yankai Jiang, Pranathi Wuppuluru, Devesh Tiwari
TL;DR
Variable datacenter demand and negative grid prices make constant SMR output costly, while reactor physics complicates load following. FissionReady tracks module-level reactor state and coordinates power and workload scheduling across two timescales. Its evaluation reports substantial water and grid-cost savings with service guarantees and no involuntary shutdowns.
Problem
Variable datacenter demand and negative-price surplus exports make it difficult for SMR-powered data centers to match generation to load efficiently, while xenon dynamics constrain safe load following.
Source: §1 · p. 1Method
FissionReady tracks per-module fuel age and iodine-xenon state to derive feasible power bounds, then coordinates reactor output, grid purchases, and batch dispatch through a two-timescale scheduler.
Source: §1 · p. 1, §4.2 · pp. 4–5Results
Takeaways & Limitations
In the reported evaluations, FissionReady preserves workload quality of service while avoiding involuntary shutdowns in SMR-powered datacenter operation.
Source: §8 · p. 9Between periodic calibrations, operating reactors may develop small transient mismatches from the reactor model; the evaluation tests their effect on shutdowns and water savings.
Source: §6 · p. 7
Abstract
from arXiv · showhide
Data centers are increasingly exploring small modular nuclear reactors (SMRs) as a carbon-free power source, but variable datacenter demand and negative grid prices require the SMR plant to load-follow rather than run at constant output. Load following is uniquely challenging and complex for SMR plants due to underlying nuclear physics. We propose FissionReady, a datacenter scheduler that tracks each module's fuel age, xenon state, and remaining flexibility to determine per-module operating envelopes, then coordinates load-following, batch deferral, and grid purchasing through a two-timescale optimization. FissionReady achieves zero involuntary shutdowns and zero batch deadline misses while reducing water consumption by roughly one-third and grid cost by roughly half compared to a same-sized base configuration.
1 Introduction
Variable datacenter demand can make an SMR plant’s fixed output wasteful, while load following is constrained by reactor physics and negative-price exports. FissionReady tracks module-specific reactor state and coordinates power, workload, and grid decisions; its evaluation reports lower grid cost and water use while preserving workload guarantees.
Background and Motivation
- Variable, bursty datacenter demand makes peak-sized SMR plants produce surplus power, increasing electricity waste, cost inefficiency, and water footprint.Exporting surplus during high renewable generation can incur costs when wholesale prices are negative.Source: §1 · p. 1
Brief Overview of FissionReady
- FissionReady tracks each module’s fuel age, xenon buildup, recent power history, and remaining flexibility to respect nonlinear xenon dynamics.Modules at the same output can have different safe futures because their reactor states differ.Source: §1 · p. 1
- A two-level planner coordinates module ramps, grid purchases, and batch deferral within physics constraints, while an inner scheduler dispatches jobs against short-term demand fluctuations.The approach maintains workload quality of service while reducing water consumption.Source: §1 · pp. 1–2
Flagship Results of FissionReady
- FissionReady cuts grid electricity cost by roughly half, reduces water consumption by 31%, and lowers SMR energy waste from 29.8% to 1.7% while preserving zero batch deadline misses.The evaluation compares against same-sized baselines and a competitive fixed load-following design.Source: §1 · p. 2
2 Background
Reactor reactivity describes distance from the critical state, and controllable maneuvering margin is only the excess left after operational and safety requirements. Xenon’s delayed response to power changes makes reactor history relevant to that margin.
Reactor Reactivity
- Reactivity, ρ, measures how far the reactor is from criticality; at k_eff = 1 the chain reaction sustains steady power.Above 1, power tends to rise; below 1, it tends to fall.Source: §2 · p. 2
Xenon Poisoning and Excess Reactivity
- Xenon-135 builds up over hour-scale dynamics, so its concentration does not respond immediately to power changes.Iodine-135 continues decaying into xenon after power is reduced, allowing xenon concentration to keep increasing for several hours.Source: §2 · p. 2
Load Following and Maneuvering Margin
- Only the reactivity remaining after safety and operational demands is available for load-following maneuvers.Shutdown margin, temperature defect, power defect, and equilibrium xenon all consume reactivity before the remaining maneuvering margin can be used.Source: §2 · p. 2
3 FissionReady: Insights and Design Drivers
A module’s safe load-following flexibility depends on both fuel-cycle position and recent power history: fuel depletion changes its available margin, while delayed iodine-xenon dynamics make identical current outputs compatible with different future trajectories.
Impact of fuel cycle of a module and staggered fueling among different modules
- In a pressurized-water SMR, available load-following margin shrinks from 5500 pcm at beginning of cycle to approximately 0 pcm at cycle end.The reduction reflects fuel depletion and increasing shutdown, temperature, and power demands.Source: §3 · pp. 2–3
- Lower capacity factor leaves a module with more available margin for changing its power level.Modules in a staggered-refueling fleet can occupy different fuel-cycle stages at the same time.Source: §3 · pp. 2–3
Impact of fuel cycle history, historic power generation level, and xenon dynamics
- A module’s compensation ceiling is the maximum xenon reactivity its control rods can offset at its current fuel age, falling from roughly 8000 pcm at BOC to 2500 pcm at EOC.A safe proposed ramp depends on whether the later xenon buildup remains within that ceiling.Source: §3 · p. 3
- Recent power history matters because iodine already in the core continues decaying into xenon after a power reduction, shaping the later xenon transient.The resulting iodine and xenon concentrations determine the future xenon trajectory; the transient depends on the interaction of fuel-cycle position and history.Source: §3 · p. 3
- After the same 50%-to-35% ramp-down, Module A peaks near 2220 pcm below its 2500-pcm ceiling, while recently ramped Module B reaches roughly 2600 pcm and exceeds it.Both are end-of-cycle modules currently at 50%; Module B had been at 65% three hours earlier.Source: §3 · p. 3, Figure 1 · p. 3
Summary
- A scheduler seeking to exploit load following must track both fuel age and recent iodine-xenon dynamics.Fuel-cycle position and capacity factor affect margin, while recent ramp history affects the xenon state and its future trajectory.Source: §3 · p. 3
4 FissionReady Scheduler Design
FissionReady converts module-specific reactor constraints into operating envelopes, then coordinates power, grid purchases, and batch work through a two-timescale scheduler. Its design combines physics-derived flexibility estimates with an optimization that respects workload and plant constraints.
4.1 FissionReady’s Reactivity Headroom Concept
- Reactivity headroom measures the gap between a module’s compensation ceiling and its predicted peak xenon at a candidate target power.The compensation ceiling is the xenon reactivity the module can tolerate; predicted peak xenon represents what it will experience.Source: §4.1 · p. 4
- FissionReady predicts peak xenon by integrating the iodine-xenon dynamics from each module’s current state, then finds the lowest power with nonnegative headroom.The current iodine and xenon concentrations encode recent power history; the minimum feasible power can be found by binary search over the predicted peak.Source: §4.1 · p. 4
4.2 FissionReady’s Optimization Approach and Architecture
- A physics subproblem solves the nonlinear reactor dynamics and converts each module’s feasible range into linear power bounds for the scheduling optimizer.The bounds are computed from headroom and a binary search for the minimum feasible power.Source: §4.2 · p. 4
- The outer planner re-solves over a rolling horizon every several inner scheduling buckets, accounting for xenon transients, batch deadlines, and grid conditions.It sets module power, grid purchases, steam bypass, and batch energy allocation; the inner scheduler dispatches jobs within the plan.Source: §4.2 · p. 5
4.3 FissionReady Scheduling Details
- The objective balances service shortfall, batch deadline misses, plant and grid resource use, and wasted energy while honoring module operating envelopes.A water-use term can favor below-demand SMR generation and grid substitution during renewable-heavy hours when module headroom permits.Source: §4.3 · p. 5
- SMR generation is penalized for water intensity and fuel cost, including when steam bypass diverts output away from the datacenter.Capital cost and fixed operations and maintenance are excluded as sunk costs over the planning horizon.Source: §4.3 · p. 5
4.4 Constraints
- The planner allocates aggregate batch energy across future steps while enforcing power balance, deadline feasibility, queue conservation, and per-module power limits.Online demand and batch jobs already dispatched contribute to nondeferrable demand; grid imports are limited by the site interconnection.Source: §4.4 · p. 5
- At each inner step, updated module states yield new headroom and operating envelopes that feed the outer optimizer alongside workload and grid information.The inner scheduler serves nondeferrable demand using SMR output and grid power, then spends the planned batch-energy budget by dispatching jobs.Source: §4.4 · pp. 5–6
5 Evaluation Methodology
The evaluation uses batch and latency-sensitive workloads, production-trace arrivals, and a fixed-size datacenter cluster. It compares three baseload-style SMR configurations against metrics covering workload service, water, and grid cost.
Workload and Evaluation setup
- The workload suite combines compute- and memory-bound batch benchmarks with four latency-sensitive TailBench services under strict SLAs.Job arrivals follow the eight-day Alibaba cluster trace; online services are served immediately, while batch jobs may wait up to 12 hours before a deadline miss.Source: §5 · p. 6
SMR plant configuration
- At this cluster scale, Alibaba-trace analysis finds mean and peak offered demand of 7 MW and 16.1 MW; six 1.7-MW modules provide 10.2 MW, sized for 90th-percentile demand under normal operation.During a scheduled refueling outage, five modules provide 8.5 MW, covering mean demand of 7 MW; the scheduler manages a grid interconnection capped at 1.7 MW.Source: §5 · p. 6
Figures of merit
- The evaluation reports workload-service, water-consumption, and grid-cost metrics against three baseload configurations sized to 7.0 MW, 10.2 MW, and 16.1 MW demand.Water includes direct SMR cooling and indirect grid-generation water; each baseline runs available modules at full power and buys grid power for shortfalls.Source: §5 · p. 6
6 Results and Analysis
FissionReady improves water, grid-cost, and SMR-efficiency outcomes over same-sized baselines while preserving workload guarantees. Year-long and sensitivity evaluations report continued operation without deadline misses or involuntary shutdowns under the tested conditions.
FissionReady’s effectiveness
- FissionReady has zero involuntary shutdowns and a 0.03% online SLA violation rate, while mean-demand Baseline-1 misses 256k batch deadlines and P90 Baseline-2 wastes excess generation.Peak-sized Baseline-3 reduces grid dependence but wastes thermal output that still consumes cooling water and fuel.Source: §6 · p. 6
FissionReady in Action
- In the 36-hour example, Module 2’s headroom falls from about 750 pcm to near zero after ramp-down, so FissionReady shifts later reductions to the less-constrained Module 1.Grid purchases during the window average −$1.5/MWh versus $28.5/MWh over the full window, and come from lower-water grid hours on average.Source: §6 · pp. 6–7
FissionReady’s effectiveness
- 31% less water and roughly half the grid cost, with SMR energy waste reduced from 29.8% to 1.7% versus Baseline-2.FissionReady also records zero batch deadline misses, 0.78 h average batch wait, and 2.32 h P99 batch wait against a 12 h deadline.Source: §6 · p. 6, Figure 2 · p. 7
Scheduling overhead
- For six modules and a 48-hour horizon, optimization overhead averages 35 ms, less than 0.006% of a 10-minute scheduling interval.Overhead depends on module count rather than cluster size.Source: §6 · p. 7
Robustness to model mismatch
- In the aged-fleet mismatch test, involuntary shutdowns occur only beyond a 15% perturbation to xenon-destruction dynamics; increasing the safety reserve beyond 300 pcm can limit or eliminate them.The perturbation makes xenon decay more slowly after power reduction, increasing the xenon peak and consuming more headroom than the scheduler expects.Source: §6 · p. 7
Why Per-Module State Tracking Matters
- FissionReady outperforms Fixed Load Following and day-ahead DCSMR at 10.2 MW sizing; Fixed Load Following causes 11 shutdowns and 212 hours of deadtime.FissionReady has zero shutdowns and deadtime; Fixed Load Following’s grid costs are 30× higher and SLA violations nearly 5× more.Source: §6 · p. 7, Table 1 · p. 8
Robustness to workload forecast error
- At 25% forecast noise, average batch wait rises from 0.78 h to 0.96 h and online SLA violations from 0.03% to 0.34%, while deadline misses and involuntary shutdowns remain zero.The inner dispatcher observes the actual queue at each step, while dispatch remains constrained by each module’s current physics envelope.Source: §6 · p. 8
Year-long operation
- Over one year, water use falls from 211 million liters to 139 million liters and SMR energy waste from 31.3% to 1.6%, with zero deadline misses or involuntary shutdowns.The fleet absorbs three staggered refueling outages; average batch wait is 0.80 h and online SLA violations are 0.07%.Source: §6 · p. 8
7 Related Work
Prior work has examined flexible nuclear operations and energy-aware workload scheduling, but fine-grained module-aware scheduling for SMR-powered data centers remains largely unexplored. FissionReady combines module-level reactor-state awareness with workload management and grid coordination.
7 Related Work
- Prior nuclear-flexibility work spans plant dynamics, reactor-level control, ramp constraints, multi-unit dispatch, and fuel-cycle-aware unit commitment.These lines of work address plant operation and power-system dispatch, while FissionReady targets datacenter operations.Source: §7 · pp. 8–9
- Fine-grained module-aware SMR datacenter schedulers remain largely unexplored; FissionReady couples per-module physics with workload management and grid use.It tracks fuel age and iodine-xenon state to compute module-specific power ranges and headroom, then coordinates fleet planning with finer-grained dispatch.Source: §7 · pp. 8–9
- FissionReady jointly accounts for batch deadlines, online-service quality of service, grid price, and water consumption in its two-timescale scheduler.The planner schedules fleet output and grid use over a horizon, while online dispatch makes finer-grained decisions.Source: §7 · p. 9
- Datacenter scheduling has used temporal and geographic workload shifting and increasingly co-optimizes water, carbon, and electricity cost across sites.Those approaches treat energy supply as externally given and do not address SMR water consumption.Source: §7 · p. 9
8 Conclusion
FissionReady combines module-level reactor physics with workload management through reactivity headroom and two-timescale control. Its main and year-long evaluations report water and grid-cost savings while preserving workload quality of service and avoiding involuntary shutdowns.
8 Conclusion
- FissionReady couples per-module reactor physics with datacenter workload management.It tracks fuel age and iodine-xenon dynamics, represents maneuvering flexibility through reactivity headroom, and combines this abstraction with two-timescale control.Source: §8 · p. 9
- The main and year-long evaluations report substantial water and grid-cost savings while preserving workload quality of service and completely avoiding involuntary shutdowns.The passage gives no numerical savings values.Source: §8 · p. 9
- The paper presents SMR-powered datacenter scheduling as a case study in unique challenges and opportunities, calling for joint consideration of physics, workload, and grid dynamics.Source: §8 · p. 9