Chiller Plant Availability: When Maintainability Cannot Buy 99.9%
A data center cooling plant has three identical 1,200 kW chillers: two running and one standby, but the standby requires a 3-minute manual changeover. Each chiller has a demonstrated MTBF of 1,200 operating hours. Corrective repair time is lognormal with mean 8 h, and the compressor is not stocked, adding a mean 24 h logistics delay. Preventive maintenance runs every 1,000 operating hours and takes 6 h; the annual regulatory overhaul adds 12 h of planned downtime. The service-level agreement promises 99.9% availability of cooling capacity, measured monthly. Last month the plant logged 10 corrective work orders and 14 preventive tasks across 12,000 chiller-operating hours, and the customer is threatening penalties. Operations proposes buying a compressor spare and adding a second technician; engineering proposes extending the PM interval to 2,000 h.
1. Compute last month's operational availability and the downtime budget implied by 99.9%. Break downtime into buckets and identify what dominates.
Read the worked answer
Downtime by cause, over 12,000 chiller-operating hours:
- Corrective repair: 10 x 8 = 80 h
- Logistics delay: 10 x 24 = 240 h
- Preventive maintenance: 14 x 6 = 84 h
- Regulatory overhaul: 12 h
- Total: 416 h
Operational availability Ao = 12,000 / (12,000 + 416) = 96.65%.
By comparison, inherent availability using repair time only is 12,000 / 12,080 = 99.34%, and achieved availability including PM but not logistics is 12,000 / 12,176 = 98.55%. The gap between 99.34% and 96.65% is entirely support and logistics, which is exactly what an availability metric is meant to expose.
The 99.9% target allows 0.001 x 12,000 = 12 h of downtime in the month. The plant consumed 416 h, about 35 times the budget.
Ranking the drivers: logistics 240 h (58%), PM 84 h (20%), repair 80 h (19%), overhaul 12 h (3%). Two conclusions follow immediately. First, no repair-time improvement can close a 35x gap on its own. Second, PM plus overhaul alone is 96 h, eight times the entire monthly budget: the maintenance program by itself already violates the SLA before a single failure occurs. Note also that this is a capacity SLA, but the calculation above is per-unit availability; the two are not the same quantity and the contract must say which is measured.
2. Evaluate the three proposals individually and in combination against the 99.9% target. Can any combination get there? What do you recommend?
Read the worked answer
Individually:
- Compressor spare removes the 240 h logistics delay: downtime 176 h, Ao = 12,000 / 12,176 = 98.55%.
- Second technician, halving mean repair time to 4 h, saves 40 h: downtime 376 h, Ao = 96.97%.
- PM interval to 2,000 h halves PM events to 7, saving 42 h: downtime 374 h, Ao = 96.97%.
All three together: 10 x 4 (repair) + 7 x 6 (PM) + 12 (overhaul) = 94 h, giving Ao = 12,000 / 12,094 = 99.22%. Still short, and that optimistic figure assumes extending PM from 1,000 h to 2,000 h does not increase the failure rate, which is doubtful on a machine with a 1,200 h MTBF: you would be scheduling maintenance after the expected failure.
The structural point: with 10 failures per 12,000 h, even instantaneous repair leaves 84 + 12 = 96 h of planned downtime, which is 99.2% at best. Adding logistics delay makes 99.9% unreachable by maintainability alone. Redundancy is the only route: an N+1 configuration with automatic start and transfer, sized so that one chiller loss still meets IT load, with the 3-minute manual changeover replaced by automatic transfer and a tested load-shedding sequence. Availability then depends on two of three units being available, and a single failure or a single PM event costs nothing.
Recommendation: buy the spare (high value, low cost), keep the second technician if repair time is also on the critical path for other assets, and reject the calendar-based PM extension in favor of condition monitoring (vibration, oil analysis). Fund automatic transfer as the primary fix, and renegotiate the SLA to state capacity, load assumptions, and whether planned downtime counts. Verify with a 90-day trial and a staged changeover test.
3. The customer wants a monthly availability report with penalties. What statistical and contractual pitfalls would you raise before signing the revised SLA?
Read the worked answer
Several, and they matter more than the arithmetic:
Sample size. Twelve thousand unit-hours is a small sample for a 99.9% claim. A month with zero failures is consistent with anything from excellent to mediocre performance; the confidence interval on availability from a single month is very wide. A monthly penalty clause will produce disputes driven by randomness rather than by engineering. Propose quarterly or rolling 12-month measurement with a defined start-up baseline.
Definition of uptime. Capacity availability, unit availability, and 'IT load served' are different metrics. If the plant sheds non-critical load during a chiller outage, a naive calculation may score it as available. Specify the load profile, the ambient design condition, and whether exceeding design temperature counts as downtime.
Planned versus unplanned. Operational availability typically includes planned downtime; many commercial SLAs exclude it. PM plus overhaul alone is 96 h per month here, so this clause determines whether the target is even approachable. State explicitly what is excluded, and cap it.
Exclusions for cause. Power failure, water quality, customer-initiated load changes, and maintenance deferred at the customer's request must be handled in advance.
Measurement and evidence. Define the data source: work-order timestamps audited, chiller run-hours from the BMS, not operator recollection. Include a joint monthly review and a dispute step before penalties apply.
Penalty symmetry. If penalties apply below 99.9%, there should be a documented gain-share or credit for sustained performance above it, otherwise the supplier carries all variance and prices it back in. Whatever is agreed, model the expected penalty against the achieved 96.65% first, so nobody signs a number that is currently unattainable.