Between July 19 and 21, an auto parts manufacturer in Panyu moved its factory 38 kilometers to a new site. The self-built server room held 38 rack servers running ERP, MES, file services and a virtualization cluster, plus 2 Synology NAS units, 1 firewall, 2 core switches and a UPS. One requirement: ERP and MES could be down no more than 4 hours, because the line restarts Sunday night.
Preparing 38 units is five times the work of moving them.
Phase one, three weeks out: inventory and planning
Asset inventory, week one. We logged all 38 servers: model, serial, U position, IP, service carried, RAID configuration and dependencies. That table became the single source of truth. It also turned up 6 machines the customer had lost track of and 3 that had been offline for months but were still drawing power.
Labels, week one. Both ends of every cable got a laminated label reading local port to remote port, and every machine got three, front, side and top, since only some faces show once boxed. Cut this corner and you pay tenfold at the new racks.
Preconditions, week two. A room is not a server room. We checked eight items:
| Item | Requirement | State of the new room |
|---|---|---|
| Power | Independent circuits at 1.5x total rack load, dual feed preferred | 3 circuits at 32A, single feed. Capacity fine, no redundancy; we recommended an ATS |
| UPS | Matched capacity, new batteries | 2 new 6kVA online units plus a battery cabinet, 40 minutes runtime |
| Cooling | 1.3x total heat load, running 24 hours | 2 x 5-HP precision units in N+1, adequate |
| Floor loading | 800-1,200 kg per square meter, racks full | First-floor room, adequate |
| WAN links | Provisioned and tested before the move | 2 x 500 Mbps lines, primary and backup, ordered two weeks early |
| Cabling | Ladder rack, tray, fiber to the desk | Installed and tested 10 days early, with a link report |
| Grounding | 4 ohms or lower | Measured 1.8 ohms, adequate |
| Fire and access | Gas suppression, badge records | Heptafluoropropane plus card access, adequate |
Network plan, weeks two and three. Keep the subnet and move it whole, here 10.10.0.0/16, changing nothing but physical location. Change it and every reference has to follow: ERP and MES clients, printers, scanners, PLCs, firewall policy, VPN, camera IPs, and missing one gives you an outage. That decision cut downtime from a possible two days to under four hours.
Backup, week three. Full export of the ERP and MES databases, the virtualization configuration, firewall policy and switch configs, to a portable drive and the cloud. No relocation starts without it.
Phase two: designing the sequence
Moving everything at once means stopping everything, so we split the work three ways:
| Batch | Equipment | Window | Impact |
|---|---|---|---|
| First | 12 non-core: file servers, test machines, backup NAS, print services | Sat 08:00-14:00 | None, nobody works Saturday |
| Second | 18 virtualization hosts, with vMotion evacuating VMs first, plus 2 NAS | Sat 15:00 to Sun 06:00 | Some VMs paused, non-production only |
| Third | 8 ERP and MES database servers, firewall, core switches, UPS | Sun 06:00-10:00 | 3 hours 40 minutes of downtime |
Core equipment moves last, but the network comes up first. On Saturday, once batch one was across, we configured the core switches and firewall, connected the dedicated lines and verified internal, external and VPN access. By Sunday that network had 16 hours of runtime behind it, so Sunday was not spent fixing the network and restoring ERP at once, which four hours would never have covered.
Phase three: pulling, packing, transport
- Shut down properly. Servers go through the OS so the RAID cache flushes back, NAS units shut down through DSM, virtualization hosts enter maintenance mode. Pulling power can corrupt RAID metadata and the file system.
- Label disks by slot and leave them in. Every drive gets its slot number, 1 through 4, and travels inside the chassis. Back in the wrong order, an array may not be recognized or may start a rebuild, and a fault during a rebuild is a data disaster. This is how moves most often fail.
- Pack for shock. Antistatic bag, pearl cotton and the original carton per server, a flight case if the carton is gone, stretch film over rack gear. Use an air-ride van: July in Guangzhou brings downpours and sun.
- Protect fiber separately. Keep a jumper's bend radius at least 10 times its diameter — about 3 cm in practice. Cap the connectors and reel it. A broken fiber is invisible, and shows up as packet loss that takes a day to trace.
- Treat UPS batteries as dangerous goods. Disconnect them, tape the terminals, keep them upright, box them apart. One short circuit in a crate is a fire; spent batteries go to a licensed recycler.
Phase four: the power-on order
Powering up follows the dependency chain, one step at a time, each verified before the next:
- UPS and distribution. Energize the UPS and confirm input voltage, charging and stable output with a multimeter, then the PDU.
- Network. Core switch, firewall, then access switches, one at a time, 2-5 minutes each, checking port lights. The network first makes management ports reachable.
- Storage. NAS, SAN and arrays need 10-20 minutes before every disk reports ready. Skip it and the servers fail to mount.
- Virtualization hosts. Wait for the cluster, then start VMs one by one. A one-click power-on creates a boot storm whose current spike can take the UPS down.
- Database and application servers. ERP and MES last, then verify consistency with DBCC CHECKDB.
- Business verification. A server that powers on is not a service that works. ERP must cut an order, MES must collect data, scanners, printers, shares and VPN must all work. We had a 32-item acceptance list and signed every line.
At 09:40 on Sunday all 32 items passed, actual downtime was 3 hours 40 minutes, and the line started normally Monday.
Three pitfalls
Disks out of order. One customer pulled every drive from 8 servers, boxed them loose, pushed them back in. Three arrays reported a foreign configuration, one rebuilt to 60% before an old disk died, and with two disks offline in RAID 5 the array was gone. Recovery cost RMB 6,800.
No external link. On moving day the new site had no dedicated line and business stopped for a day and a half. Provisioning takes 5-15 working days, so the order goes in three weeks ahead.
Invisible conditions. Forty-two units put out about 25kW, and two 3-HP cabinet units will not hold that; the room hits 40 degrees C and every server throttles. Floor loading and grounding bite the same way. Measure, do not assume.
What a relocation costs
| Item | Scope | Amount |
|---|---|---|
| Survey and design | Both rooms surveyed, asset inventory, network plan, batching, acceptance list | RMB 6,000 |
| Base construction | Cabling, racks, cable tray, UPS installation, materials separate | RMB 28,000 |
| De-rack, pack, transport | 38 units plus accessories, air-ride van over 38 km, loading | RMB 9,500 |
| Re-rack and recovery | Cable management, commissioning, system restore, integration testing, acceptance | RMB 18,000 |
| On-site support | 72 hours on site plus 3 months remote | RMB 6,000 |
| Total | RMB 67,500 |
Moving it yourself saves this fee, but one data loss costs RMB 3,000-20,000 to recover and one lost day costs a manufacturer RMB 50,000-200,000. The value is not the hauling. It is nothing going wrong.
ACCPC provides turnkey relocation and server room construction across Guangzhou and the Pearl River Delta: survey, cabling, migration, recovery, acceptance report and support. We also coordinate IDC colocation with a licensed provider offering dual feeds, UPS, diesel, 24-hour staffing and BGP multi-homing. Call 020-39029800.