FAQ

Yes. We support night and weekend cutover windows to minimize business impact, with a rollback plan prepared in advance.

For production databases on minicomputers, external SAN storage is recommended for better scalability and data reliability; test environments can use internal disks.

For large-scale LLM training, prefer InfiniBand: lower latency and higher bandwidth. For small and medium clusters on a budget, Ethernet RoCE cuts hardware cost and is easier to operate.

Hardware faults: hardware alerts, abnormal reboots, dropped accelerator cards, memory or drive errors. Software faults: SQL exceptions and database process crashes with logs free of hardware errors.

Just basic equipment details: brand and model, quantity, region and expected service level - the more detail, the more accurate the quote. Submit via the website inquiry form, phone or WeCom. ZhongLianDa promises a preliminary quote within 2 hours (1 business day for complex projects). The proposal covers service scope, response SLA, coverage and a full cost breakdown with no hidden terms, and free revisions until you are satisfied.

Yes, within a unified cutover window, but the risk is higher. We recommend implementing them separately to limit the impact scope of any failure.

The standard migration service includes data comparison, object validation and business verification to guarantee pre- and post-migration consistency.

It depends on the failure type. Logical damage (accidental deletion, partition or file system failure) recovers faster: a few hours for simple cases, 1-3 days for complex ones, with over 90% success. Physical damage (firmware corruption, head failure, platter scratches) needs specialized equipment, usually 3-7 days, with a 60%-80% success rate. ZhongLianDa's data recovery center has a dust-free lab and professional equipment; we test free of charge first, quote after recoverability is confirmed, and charge nothing if recovery fails.

Yes. We provide one-stop service from racking, OS installation and cluster deployment to data migration, covering Oracle database migration, RAC setup and GPU computing platform deployment. Migration emphasizes data safety and business continuity, with planned maintenance windows.

Yes. The correct order is: storage first (FC switches, storage arrays), then network devices (core switches, firewalls), then database servers, and finally application servers. This prevents applications from starting without storage and databases from losing network paths, reducing the risk of data corruption. ZhongLianDa prepares a standardized power-outage restart SOP document for customers - free for contracted customers.

Hardware inventory, power-on identification, stress testing, network connectivity, power and temperature telemetry, error-free logs, delivery documentation and training.

It depends on the contract. Basic racking and power-on can be bundled; complex cluster networking and distributed training environment deployment are usually billed as a separate implementation service.

Yes. We provide business licenses, case studies, authorization letters and maintenance plans as required by the tender, and assist with Q&A during bidding.

Assembled: lower cost and flexible hardware choices; drawbacks are compatibility risk, no OEM warranty and harder fault diagnosis. Branded: factory-tested, full technical documentation and OEM support; drawbacks are higher price and limited configuration changes.

Choose U.2 NVMe for fast model loading and dataset reads; for ordinary logs and cold storage, SAS drives offer better cost efficiency.

Inlet temperature 18-24 C, humidity 40%-60%; avoid condensation and dust. Sustained temperatures above 28 C cause repeated GPU throttling and shorten GPU lifespan.

Power: redundant high-wattage 12V output able to handle transient peak loads. Cooling: high-static-pressure fans with airflow directed at the GPUs; multi-GPU systems must keep GPU temperature spread controlled to avoid thermal throttling.

Self-assembly is possible, but multi-GPU platforms demand strict compatibility across motherboards, power delivery, airflow and backplanes. OEM systems are stress-tested across the whole chain and are more stable, making them the better choice for production clusters.

In-stock refurbished or assembled servers: 1-3 days. New branded systems: typically 4-12 weeks, depending on GPU supply.

Create sub-users under Identity & Security → IAM, grant them the minimum sufficient permissions, and avoid logging into the console with the root administrator account for daily work.

Free accounts have no human technical support channel. Most issues can only be resolved via online research or by rebuilding the instance. Only paid accounts can submit support tickets.

Use top or htop to view processes, identify the high-usage program and stop unneeded ones. Free-tier hardware is limited — over-provisioned programs will max out the CPU.

Boot volume backup is Oracle-managed backup; a snapshot is a user-triggered manual snapshot. Both consume storage space.

Snapshots consume storage capacity and count toward the 200 GB free disk allowance — excess is charged. Snapshots are well suited for data backup.

Free instances have limited disk performance; paid instances can raise the block storage performance tier. Avoid running high-IO workloads on free machines.

Detaching a volume does not delete anything — data stays intact on the volume. Only manually deleting the volume erases it.

Console expansion only enlarges the block storage capacity; inside the OS you must also grow the partition and the file system, otherwise the new space is not visible.

Possible causes: server resources exhausted, firewall rules blocking, or free-tier risk control. Log in via the serial console first to check system load.

Log in through the serial console; you can change the root password directly without reinstalling the instance.

Several failed key attempts within a short window made the server refuse the connection. Wait a while and retry, or clean up extra private keys on your local machine.