Facility Piping & Pressure-System Mechanical Engineering

The moment a hall is plumbed for direct-to-chip liquid (Chapter 5.4), it has a pressure system: charged pipe carrying coolant through a building full of energized racks, through twenty years of thermal cycles and the occasional pump trip. Mechanical codes, surge physics, pump hydraulics, corrosion electrochemistry and structural restraint govern it, not ASHRAE setpoints. These are mature disciplines from process plants and building services. AI density puts them in a setting they were not written for: water routed close to energized busbars, velocities that can erode a passage not sized for them, and GPU-hours lost whenever a leak or a tripped pump takes a row offline.

Build to the wrong code section and your weld QA and hydrotest regime are either too heavy or, worse, too light for the duty. Skip the pump-trip and isolation surge model and a routine power blip can burst a manifold over a powered rack if the transient crosses a component’s pressure limit, so establish both maximum and minimum pressure before you buy the circuit. Mismatch the pump to the system curve and you either cavitate at the low-pressure node or dead-head into a throttled, wasteful operating point. Couple an aluminum manifold electrically to a large copper header through an incompatible wetted circuit and the galvanic cell can eat the small, less-noble aluminum area on a schedule you did not budget for. Trace the steel CDU, the bonds and the electrolyte paths as part of that circuit: a bigger pump cannot repair a corrosion mechanism left open at procurement.

The first irreversible decision is jurisdictional: what code do the charged loops live under? In North American projects whose adopted basis invokes ASME B31, the Code for Pressure Piping, the live fork is which section, and you settle it before fabrication. For closed-loop chilled-water and technology-cooling-system (TCS) loops carrying water or water-glycol at building-services pressures and temperatures, B31.9 (Building Services Piping) is a candidate, but only inside its scope and limits. B31.3 (Process Piping) is a separate candidate where its scope and fluid-service requirements apply. Higher pressure or an owner preference alone does not make a section govern. Record the adopted edition, service classification, exclusions, and examination and test basis, and compare the actual requirements before claiming that one path is lighter.

In the EU/EEA the framework changes entirely, not just its number. The Pressure Equipment Directive (PED, 2014/68/EU) starts its scope test at PS > 0.5 bar, after exclusions. Article 4 sets the essential-safety-requirement thresholds. Article 13 and Annex II classify pressure equipment by PS, volume or DN, fluid group and phase, and Article 14(6) governs global conformity assessment of assemblies. Article 4(1)/(2) equipment and assemblies follow conformity assessment and CE marking. Article 4(3) equipment and assemblies follow sound engineering practice and must not bear the PED CE mark. EN 13480 (Metallic Industrial Piping) offers a design route for metallic industrial piping. Check the applicable edition and its current harmonized listing, because applying a standard does not by itself complete PED conformity assessment.

The practical consequence for a global operator is that a reference design validated to B31.9 in Virginia cannot simply be shipped to Frankfurt or Dublin. The EU loop needs a PED scope and exclusions test, equipment classification under Article 13 and Annex II, and assembly assessment under Article 14(6). The actual equipment categories and assembly procedure set the Notified Body involvement and the CE documentation. Resolve this before you freeze the global skid, because one jurisdiction’s code does not automatically satisfy another, and the cost recurs with every region you standardize.

Steady-state pressure is the easy part, because every loop is designed to a static pressure rating. The hazard is the transient: the pressure spike when a velocity changes suddenly. Its first-order size is the Joukowsky surge, ΔP = ρ·a·ΔV, where ρ is fluid density, a is the pressure-wave speed in the fluid-filled pipe and ΔV is the velocity change. Screen a at roughly 900–1,500 m/s in water-filled metal pipe (lower in flexible hose), then calculate the actual celerity for your fluid, pipe, wall, supports and geometry. A 2 m/s instantaneous stop in water near 1,000 kg/m³ screens at roughly 18–30 bar above operating pressure across that range. Before using the screen as a design pressure, compare valve stroke time with wave travel time and the real network boundaries. In a data hall the spike lands on manifolds, quick-disconnects and cold plates sitting directly above energized GPUs.

Two events drive surges, valve closure and pump trip, and they behave oppositely. A fast valve closure sends a positive (compression) wave upstream toward the source. A pump trip sends a negative (rarefaction) wave downstream from the discharge, because fluid momentum carries on without the driving head. That happens on a utility blip, ATS transfer or breaker event where power or inertia ride-through is absent. The negative wave is the dangerous one. If local pressure falls to the coolant’s vapor pressure, the column separates (a vapor cavity forms), and when the cavity collapses the rejoining columns slam together and can produce a surge that exceeds the Joukowsky estimate. AFT lists column-separation collapse as one of four ways the real peak can exceed the equation. The others are a change in pipe size (apply the equation to the pipe with the largest ΔV), several surges arriving together and adding, and line packing in long or high-friction runs, where flow does not stop with the first wave. A hand calculation is therefore a screen, not a design tool.

That leaves the design question: do you model the credible pump-trip and valve-isolation events with a transient method such as the method of characteristics, or trust a static rating with a fudge factor that never establishes the maximum and minimum pressure? For any non-trivial charged loop in a dense hall, the answer is the model, using software in the lineage of AFT Impulse and equivalents. Run it across the credible event set: single and simultaneous pump trips, the worst-case valve slam, and an emergency stop that drops all pumps at once. It gives the maximum and minimum pressure at every node, including the negative-pressure nodes where column separation and check-valve slam live, which a steady-state rating never reveals.

The pump is the heart of every charged loop, and two failures recur. The first is cavitation. It starts when local absolute pressure, usually at the impeller inlet, drops below the coolant’s vapor pressure. Vapor bubbles form, then collapse where pressure rises again inside the impeller. They pit the metal, collapse the head and roar like gravel in the casing. The guardrail is the NPSH inequality: NPSH-available (set by the loop: suction static pressure, fluid temperature and vapor pressure, suction-side losses) must exceed NPSH-required (a property of the pump at its flow) by an application-specific margin at every required operating point. Use the manufacturer’s NPSHR basis and the ANSI/HI 9.6.1 margin you select, not a universal added head. In a closed, pressurized data-center loop NPSHa is usually engineered with a generous expansion-tank and fill pressure, but check the margin across every credible flow, coolant temperature, fill pressure and redundancy state, including maximum flow with hot coolant and degraded parallel operation.

The second failure is a pump/system mismatch. A centrifugal pump delivers head that falls as flow rises (the pump curve). The loop demands head that rises roughly with the square of flow, dominated by friction (the system curve). The pump runs where the two cross: the operating point. Pick a pump that lands far left of its best-efficiency point (BEP) and you waste energy, run hot and invite recirculation and shaft load. Land far right and you risk cavitation and motor overload. The operating point also moves: a CDU loop has variable flow as racks throttle and isolate, and the system curve shifts every time a quick-disconnect is mated or a row is valved out. Constant-speed pumps chase this by throttling, which wastes energy. Variable-speed (VFD) pumps with differential-pressure control ride the family of system curves and can follow the secondary-loop duty with less throttling loss. The control loop interacts with thermal-control transients and can hunt, which is why it couples to Chapter 5.12 (controls stability) and is set during commissioning in Chapter 13.5.

Decision: choose the CDU against external circuit loss, then verify total pump duty and suction margin. The inherited flow is 320/(4.18 × 10.0) = 7.66 L/s for the zone, or 115 L/min per rack. With v = Q/(πD²/4), the common header runs at 2.31 m/s and each branch at 2.38 m/s. Both exceed the hot-water copper guidance later in this chapter, so the case assumes pipe rated for that velocity, such as stainless. Darcy–Weisbach, Δp = (fL/D + K)ρv²/2, gives 36.4 kPa for the common path and 66.9 kPa for the long branch. Add the common path to one parallel rack path: 36.4 + 66.9 + 70.0 + 30.0 + 10.0 = 213 kPa external head. Adding all four branch losses would count parallel paths in series.

The short branch’s pipe-and-fitting loss is 44.8 kPa, so its balance valve must consume 32.1 kPa to match the long path’s 66.9 + 10.0 kPa. Adding 25.0 kPa of internal CDU loss gives 238 kPa total pump differential. Pump electrical input is QΔp/η = 0.00766 m³/s × 238,000 Pa / 0.700 = 2.61 kW, using unrounded operands. That fits inside the separate 3.00 kW heat allowance in Chapter 5.6. With x = Q/Qdesign, the assumed system is Δpsystem = 238x² kPa and the pump is Δppump = 300 − (300 − 238)x² kPa. They intersect at x = 1, the computed 459 L/min duty. A real selection must supply its measured curve, efficiency, stable operating range and dirty-filter performance.

NPSHA = (pabs − pv)/(ρg) + v²/(2g) = (300,000 − 9,600)/(1,000 × 9.81) + 2.31²/(2 × 9.81) = 29.9 m, well above the assumed 3.00 + 1.00 = 4.00 m requirement. Static lift is not added again to a closed-loop friction duty, and an established suction pressure is not charged the same elevation or suction loss twice. This screens the assumed duty only; repeat it for the hottest fluid, minimum fill, transients and each pump combination.

Select the 240 kPa-external-head teaching CDU in Chapter 5.6; the 200 kPa unit misses this circuit. Flip the pipe: an 80.0 mm common header drops its loss to 14.2 kPa and external demand to 191 kPa, which brings the 200 kPa unit back into play at the cost of more pipe space and installation. Flip the fill: at 30.0 kPa absolute suction, NPSHA falls to 2.35 m and fails. The exact pressure crossover is pv + ρg(4.00 m) − ρv²/2 = 46.2 kPa absolute at this flow, and operating reserve must sit above it. The Hydraulic Institute’s explanation of ANSI/HI 9.6.1-2024 defines available and required NPSH and application-specific margin. This chapter owns the hydraulic method; Chapter 13.5 owns installed acceptance.

A charged loop is an electrochemical cell waiting to be completed. Galvanic corrosion occurs when two dissimilar metals are electrically connected in a conductive fluid. The less-noble metal (the anode) corrodes preferentially to protect the more-noble one (the cathode). Severity depends on the electrolyte, the surface state and the cathode-to-anode area ratio: a large cathode connected to a small anode concentrates the attack. Data-center loops are full of dissimilar couples: aluminum cold-plate manifolds, copper cold plates and headers, stainless CDU internals, carbon-steel facility piping and brass fittings. The worst geometry is a small anode on a large cathode, for example an aluminum fitting feeding a large copper header, which concentrates the whole galvanic current on a small area and perforates it fast. There are three defenses: (1) keep the wetted-metal set as compatible as possible, (2) trace the complete electrical and electrolyte paths before choosing dielectric unions or isolation kits, and (3) keep inhibitors, pH and the rest of the coolant chemistry inside the qualified formulation’s limits in Chapter 5.7. Retain required protective bonding, since a bond or structural connection can bridge the very joint meant to interrupt galvanic current.

Erosion-corrosion is the velocity-driven cousin. Fast or turbulent flow strips the protective oxide film faster than it can re-form, exposing fresh metal in a self-accelerating loop. The worst damage is at elbows, tees, reducers and the back of partially throttled valves, where flow impinges. Copper is notoriously velocity-sensitive, and the high flow rates of dense racks collide with its limits. Copper Development Association plumbing guidance sets maximum copper velocity near 1.5–2.4 m/s for cold water, 1.2–1.5 m/s for hot water under 60 °C, and 0.6–0.9 m/s over 60 °C. Coolant in a closed, inhibited loop is kinder to copper than potable water, but your tube and fitting supplier still sets the velocity rating; no published copper limit stands in for it.

Rack drops push harder. OCP’s Open Data Center specification, rev 0.7.0 (effective 2026-08-27), sizes them at 3 m/s in its table, faster than every copper figure above. Before you adopt it, find out what velocity your drop material is rated for. There are two ways out: upsize the drop itself, or any run that exceeds its rating, so the same flow moves slower, or switch to a metal that tolerates erosion better, such as stainless, and accept the cost. Either way, check the material and fluid against the supplier’s rated limits and against localized impingement. Ignore erosion-corrosion and the first symptom may be a pinhole leak over a rack. A new material also changes joining, chemistry and inspection requirements.

Charged pipe is heavy and it moves. Thermal expansion is the routine driver. A long warm-water header heats and grows on every load swing, and if it is rigidly anchored at both ends without relief, the thermal stress goes into the pipe, the welds and the equipment nozzles. That is how a CDU connection or a flange gasket fails early in its cycle life instead of lasting the design life. The classical solution is a deliberate flexibility scheme: expansion loops, expansion joints (bellows), anchors and guides laid out so the pipe grows in a controlled direction and is restrained against everything else. This is a formal pipe-stress problem, the same CAESAR II-class analysis the process world runs, and a dense hall complicates it with tight pipe racks, CDU galleries and the need to keep growth away from busbars and trays.

The second driver is weight and dynamics. Water-filled pipe is far heavier than the empty pipe a structural model sometimes assumes. A large header full of coolant is a significant distributed load, and a whole plumbed and charged hall adds a coolant mass the slab and supports must carry. That is the handoff to the structural basis: Chapter 6.2 owns the floor loading for plumbed, charged, energized racks and their distribution piping, and the piping engineer must feed it wet weights and dynamic (surge, seismic) loads rather than treating the slab as a given. Where a project follows OCP’s Open Data Center specification, rev 0.7.0 (effective 2026-08-27), its weight basis is a 2027–2028 design target, not a measured rack: 8,000 lb for the heaviest single rack (28 × 48 in) or 15,000 lb for an ML rack plus sidecar (56 × 48 in), with handling equipment in the moving load. Chapter 4.7 covers the rack itself. The specification also calls for wall and row manifolds sized for the whole row. Its 2027–2028 design targets, not measured deployments, are a 25 kW/ft blended average, which requires both 15 kW/ft of air-cooling capability and 25 kW/ft of liquid capacity, and a 350 kW/ft peak IT zone (335 kW/ft liquid plus 15 kW/ft air in ML zones), with 2027 peak racks expected at 125 kW/ft.

The third driver is seismic. In any meaningful seismic zone, charged pipe needs transverse and longitudinal bracing, flexible connectors at equipment, and slack at building seismic joints. A rigidly run, unbraced charged line is exactly what ruptures in an event and floods a hall of live racks. Anchoring, guides, expansion provision and seismic restraint are one coordinated system, designed to a stress basis, not eyeballed off a hanger schedule.

Decision: provide expansion acceptance and a separate overpressure path before filling the loop. Fluid growth is 1,000 L × 0.0200 = 20.0 L. Isothermal gas volume is Vgas = Vvessel × pprecharge/pabsolute, so acceptance between cold and hot conditions is Vvessel × (2.50/3.00 − 2.50/4.00) = 0.208Vvessel. The required ideal vessel volume is 20.0/0.208333… = 96.0 L. The 120 L candidate accepts 25.0 L and clears this ideal screen by 5.00 L.

Hot static pressure is 4.00 − 1.00 = 3.00 bar gauge. Adding the computed 2.38 bar operating differential gives 5.38 bar gauge at the assumed worst node, which leaves 0.617 bar below the 6.00 bar component limit. At shutoff, 3.00 + 3.00 = 6.00 bar gauge, so the circuit has no pressure reserve before transient or relief-setting tolerances. Take the candidate vessel forward for further checks, but hold the pressure-system release until the shutoff, surge and relief design clears every component. Expansion acceptance does not protect an isolated heated pocket from overpressure.

Flip the vessel: its exact expansion crossover is 25.0/1,000 = 0.0250, or 2.50%, and above that growth the candidate fails even the ideal model. Lowering hot static pressure below 4.00 bar absolute, or raising the rated component limit, opens shutoff reserve but changes the NPSH or equipment selection and needs recalculation. The ideal-gas balance is this guide’s own derivation. The vessel supplier provides the real acceptance curve, and the governing piping code, such as ASME B31.9 within its scope, sets the relief and test requirements. Chapter 13.5 tests the full pressure range at acceptance.

The code basis chosen at the top of this chapter pays off here, at acceptance. Charged piping is only as good as its joints, and three layers prove them. Weld QA requires qualified procedures (WPS), qualified welders and procedure-qualification records (PQR) under the governing code, B31.x or EN 13480, so every weld is reproducible and traceable. Non-destructive examination (NDE) checks the welds that were made. Visual inspection, radiography (RT), ultrasonics (UT) or another method is chosen, with its extent, acceptance criteria and traceability, from the adopted code, the classified service and the inspection plan. More required examination means more access and records to procure, and a PED category alone does not set a universal weld percentage.

Not every joint is a weld, and each joining method fails in its own way. Welded carbon or stainless steel suits large headers and high-consequence runs, and its defects are what the qualification and NDE above catch. Roll-grooved mechanical couplings are common on chilled-water and facility-water mains because they go in fast with no hot work. Their weak points are the gasket, whose elastomer grade must suit the glycol and inhibitor package, and thrust. The coupling's key holds the joint together, but under pressure each flexible coupling opens to the end of its gap, so a long run lengthens and pushes on its elbows and branches. Long runs therefore need rigid couplings, or anchors and guides sized for pressure thrust and surge thrust together. Press fittings on small copper or stainless lines are quicker still, but a fitting someone forgot to press can pass the pressure test and then fail in service. Specify fittings designed to leak visibly if left unpressed. Check for leaks at low pressure, within the maker's stated detection range, before raising the system to full test pressure, where an unpressed joint can still hold. Flanged joints at equipment and valves live or die on gasket choice and bolt-up sequence. Thermoplastic pipe (fusion-welded PP-R or HDPE, solvent-welded PVC-C) is appearing in secondary loops. It does not corrode, but its pressure rating falls as temperature rises, and for the same temperature swing, plain grades grow 7 to 13 times as far as carbon steel. Fiber-reinforced PP-R expands far less, about three times as much as steel, which cuts the number of expansion loops a run needs. Its fusion joints need their own procedure and trained operators.

The final gate is the pressure test. Hydrostatic testing fills the system with the approved test liquid and applies the code-calculated test pressure. ASME B31.3 (para. 345.4.2) sets the hydrostatic test at no less than 1.5 times design pressure, raised by the ratio of allowable stress at test temperature to that at design temperature. Its pneumatic test (para. 345.5.4) is 1.1 times design pressure, and the NIH Office of Research Facilities reports that B31.9 sets its hydrostatic test at no less than 1.5 times. Confirm the clause in the edition you adopt, and isolate any component rated below the test pressure. Hydrostatic testing is the default because water is nearly incompressible, so its stored energy is lower than a comparable gas fill. Trapped gas, elastic strain, projectiles and high-pressure jets still need an engineered exclusion and restraint plan. Pneumatic testing, with air or inert gas, is used where the system cannot tolerate water or cannot be fully drained. It is far more hazardous, because compressed gas stores enormous energy, so pneumatic tests demand exclusion zones, stepped pressurization and explicit engineering justification. The choice is about safety, not convenience. The pressure test also ends fabrication but starts commissioning: the cleanliness, flush, fill and chemistry steps that turn a hydrotested but dirty loop into a coolant-ready, leak-monitored, accepted system are covered in Chapter 13.5 (cooling acceptance) and tie back to the leak-detection and water-quality monitoring in Chapter 5.11.

Choose pipe diameter, pump duty, fill pressure and expansion provision as one pressure-system design. A smaller header spends head and electricity; extra fill improves suction margin but uses pressure allowance. Release the circuit only after the required steady and transient states clear every component.