RDMA Interview Questions: Queue Pairs, Memory Registration, Ordering, and Failures

RDMA interview questions focus on the boundaries that make remote memory access fast and easy to misuse: queue pairs, completion queues, registered memory, work requests, keys, ordering, and recovery. The candidate should explain what the NIC is allowed to access, how a work request becomes a completion, and which guarantees are scoped to a queue pair or operation. Official RDMA programming documentation is more useful than memorizing verbs because it exposes the setup and failure assumptions hidden by a high-level library.

Quick answer

Question	What the evidence supports	Practical move
What is a queue pair?	A queue pair combines send and receive work queues that the application posts to and the NIC processes.	Explain states, ownership, and completion handling.
Why register memory?	The adapter needs an approved virtual-to-physical mapping and access keys for DMA or remote access.	State lifetime, permissions, and invalidation.
What is a completion queue?	It reports completed work requests and status to the application.	Explain polling, moderation, and overflow.
What does ordering mean?	Ordering is scoped to operations and transport rules, not a universal global sequence.	Name the queue, connection, and operation.
How do you debug failure?	Check QP state, completions, keys, MTU, transport, and peer liveness.	Separate setup failure from data-path failure.

Read this article in three layers. An official provider or employer source establishes what the product or process says. A candidate report can suggest a useful preparation drill, but it does not prove that every candidate sees the same configuration. The final layer is an inference: a tactic that follows from the evidence. Keeping those layers separate prevents a familiar platform name from turning into an invented universal rule.

Queue pairs and state

A queue pair has send and receive queues and moves through states such as initialization, ready, and error. The application posts work requests; the NIC consumes them subject to queue capacity and transport rules. An interviewer wants to hear who owns each buffer, when it can be reused, and how a completion is matched to a request. A design that only says 'call RDMA write' has skipped the lifecycle that makes the operation safe.

Memory registration and keys

Registered memory gives the adapter the mapping and permissions needed for DMA. Local and remote keys control which operations are allowed; a remote write requires the peer to receive the address and key through an application protocol. Registration has cost and lifetime implications, so high-throughput systems use pools or long-lived regions while controlling fragmentation and access. Invalidate keys before returning a buffer to an unrelated owner. The protection model is part of the data contract.

Completions and progress

A completion queue entry reports success or an error for a work request. Polling can reduce latency but consumes CPU; interrupts or moderation can improve efficiency but add delay. The application must handle queue exhaustion, completion queue overflow, and the possibility that a signaled completion represents earlier unsignaled work. State how requests are indexed, how buffers remain live until completion, and how a stalled peer is detected.

Ordering and visibility

Reliable connected transports provide ordering rules for operations on a queue pair, but a remote write becoming visible to a consumer also depends on the synchronization protocol and memory model. Separate data movement from notification: write the payload, then publish a flag or send a message under a documented ordering rule. Do not claim global ordering across queue pairs or hosts without a coordination mechanism. A correct answer names the scope of every guarantee.

Failure modes

A failed completion, a QP entering error, a missing receive buffer, a bad key, or a link event needs a recovery path. Decide whether the application can reconnect, replay an idempotent operation, or fail the shard. Preserve the request ID and peer epoch so a delayed completion cannot corrupt a new session. RDMA's speed makes failures less visible at the application layer; instrument completions, retries, queue depth, and link health rather than waiting for a timeout.

Performance trade-offs

Measure message size, inline thresholds, batching, polling budget, registration cache hit rate, PCIe or fabric bandwidth, and tail latency. A larger batch can improve throughput while increasing queueing delay. A zero-copy path can save CPU but complicate ownership and lifetime. The best interview answer starts with a workload target and then explains which RDMA feature pays for itself and which failure complexity it adds.

A focused preparation and recovery plan

Step	What to do	Evidence to keep
1. Confirm	Read the invitation, role page, or assessment landing page. Record the deadline, timezone, timer, allowed tools, identity checks, and accommodation contact.	A screenshot or saved message with private details redacted.
2. Baseline	Complete one representative exercise without changing the rules. Measure correctness, time, and the point where you became uncertain.	The prompt, your approach, and a short error log.
3. Diagnose	Separate a content problem from a platform, network, account, or evidence problem. Do not refresh repeatedly while a timer is running.	Timestamped notes, browser version, and visible error text.
4. Rehearse	Repeat the narrow skill this article identifies. Add one adversarial case and one explanation of what would change at larger scale.	A corrected solution or one-page decision record.
5. Escalate	If the official owner must change a deadline, reopen a session, or explain a report, send a concise factual request.	The original invite, incident timeline, and requested remedy.

Frequently asked questions

Why must RDMA memory be registered?

The adapter needs an approved mapping and access permissions for DMA or remote operations.

What is the difference between a QP and a CQ?

The QP carries posted send/receive work; the CQ reports completed work and status.

Does RDMA write notify the receiver?

Not by itself in the general case; the design needs an explicit notification or synchronization protocol.

How do you recover a failed QP?

Detect the error, stop reusing buffers, preserve request identity, and reconnect or fail over under an explicit epoch.

What performance number matters?

Choose from throughput, message rate, CPU, p99 latency, and recovery time based on the workload.

Before you start

Turn the article into a small decision record before you act. Write the exact question you need to answer, the source that could answer it, and the consequence of being wrong. For an assessment, that might be whether the timer stops, whether a second display is allowed, or whether a previous result can be reused. For a technical interview, it might be whether the system promises per-key ordering, whether the verifier can prove a loop is bounded, or whether a cache update can be replayed. This framing keeps preparation tied to the target rather than to a list of fashionable keywords.

Next, separate hard constraints from choices. A deadline, access rule, memory limit, privacy boundary, or source-of-truth requirement is a constraint. A page size, data structure, model pair, retry policy, or practice order is a choice that you should justify. Write one default choice, one alternative, and the signal that would make you switch. Interviewers and recruiters can then see judgment instead of a memorized answer.

Finally, decide what evidence will close the loop. Save a confirmation, a source position, a benchmark row, a test result, a metric, or a recovery ticket. Do not collect private assessment content or credentials. The evidence should be sufficient for the next person to reproduce the decision without exposing information they are not authorized to see. If the evidence is missing, label the conclusion provisional and ask the owner for the exact next action. That discipline is useful whether the page is about a hiring platform, a production system, or a candidate-reported study.

A final quality check is simple: remove any claim that depends on an unnamed company policy, an undated forum anecdote, or a number whose denominator you cannot explain. Replace it with the narrower fact, the verification step, or the limitation. Readers usually need a safe next move more than a confident-sounding prediction. That editorial boundary also makes the draft easier to update when a platform changes its controls or a new hiring cycle starts. Use the same discipline for every source and every handoff.

How to decide what to trust

The live invitation and the instructions shown inside the assessment outrank a generic article. The provider's current help center explains default product behavior, but employers can choose different question types, timers, proctoring controls, report fields, or deadlines. An official employer page establishes the published role or event context; it does not establish every team's interview sequence. Candidate reports are useful when they identify the role, date, and product, yet they remain samples. When two sources disagree, record the conflict instead of averaging it into a made-up number. If a technical incident is involved, preserve the state first, then use the platform's support route or the recruiter who owns the invitation. This approach is slower than guessing, but it gives the reader an action that can be defended and repeated.

Final takeaway

Explain RDMA through ownership and guarantees: registered memory, queue-pair state, completion progress, scoped ordering, and a recovery path for the first failed operation.

Sources and Further Reading

• RDMA Aware Networks Programming User Manual
• Linux RDMA user verbs
• rdma_cm manual
• linux-rdma rdma-core

Research note: Sources checked 2026-09-06. Platform settings, recruiting schedules, and technical documentation can change; verify the exact invitation or current release before acting.