QoS Classes, Priority Arbitration, and Starvation Boundaries: Debug Playbook

Debug playbook

Debug Playbook for QoS Classes, Priority Arbitration, and Starvation Boundaries focuses on Per-class latency SLA compliance (real-time, interactive, best-effort) and fairness index under stress traffic.. The purpose is to turn memory observations into mechanism-backed actions with explicit owners and release-safe validation.

DRAM debug should narrow from broad symptom to one dominant mechanism. Avoid mixed-knob sweeps that produce accidental wins without causal confidence.

1. Freeze workload seed, firmware image, timing profile, and thermal setup.

2. Find first failing transition in command timeline.

3. Classify mechanism: locality loss, legality pressure, queue policy, margin drift, or RAS behavior.

4. Build focused reproducer for top hypothesis.

5. Apply minimal reversible fix and define rollback gate.

6. Re-run full performance + reliability matrix.

Debug decision tree

ROOT CAUSE TREE - QoS Classes, Priority Arbitration, and Starvation Boundaries

Per-class latency SLA compliance (real-time, interactive, best-effort) and fairness index under stress traffic. regressed
        |
reproducible with fixed seed?
      /               \
    no                 yes
    |                   |
testbench noise    localize bottleneck
                    /              \
               command path       data path
                 |                  |
             scheduler/FSM      PHY/timing/noise
                 |                  |
             timing limits      training/calibration

Stop at first failing mechanism, then patch and re-measure.

Review memo template

DRAM REVIEW MEMO - Memory Controller Scheduling / QoS Classes, Priority Arbitration, and Starvation Boundaries

1. Symptom
   - Watched metric: Per-class latency SLA compliance (real-time, interactive, best-effort) and fairness index under stress traffic.
   - Failing traffic slice: <workload/phase/class>
   - First failing transition: <row-hit/row-conflict/turnaround/refresh/training>
   - Revision tags: <firmware/controller/timing/board/package>

2. Mechanism hypothesis
   - Primary mechanism: QoS-aware arbitration overlays policy on top of raw efficiency scheduling so critical clients (for example CPU demand fetches, display, or real-time accelerators) get bounded service even when background traffic is heavy. The controller typically uses weighted priority, aging, credit/token buckets, or deadline-aware boosts to pick among ready requests. Pure fixed priority can satisfy critical latency but often starves low-priority flows; pure fairness can miss hard deadlines. Practical designs combine tiers: first enforce hard constraints (deadline/critical window), then apply weighted fairness among remaining contenders, with aging to guarantee eventual service. Arbitration decisions must be synchronized with read/write batching, bus turnaround penalties, and bank availability, otherwise QoS policy can look correct at request level yet fail at command-level execution. End-to-end QoS therefore requires both scheduler logic and upstream traffic shaping: if NoC or cache eviction policy injects pathological bursts, controller-only fixes may be insufficient. Robust implementations validate SLA behavior using adversarial traffic mixes and explicitly monitor tail latency excursions, not just average service rate.
   - Competing hypotheses: <mapping, scheduling, PHY margin, SI/PI, reliability policy>
   - Missing evidence: <command trace, queue snapshot, lane margins, CE/UE logs>

3. Proposed action
   - Smallest reversible change: <policy/register/firmware/flow>
   - Expected movement: <p99 latency, effective bandwidth, stability>
   - Regression risk: fairness, thermal drift, training robustness, field reliability

4. Signoff
   - Re-run artifact: QoS compliance dashboard with per-class SLA miss counters, arbitration decision logs, and starvation watchdog events.
   - Required owners: memory controller architect, SoC QoS architect, NoC owner, firmware power/performance owner, silicon validation owner
   - Final decision: ship, bounded rollout, rollback, or escalate

DRAM deep dive

Controller policy decides whether DRAM serves locality, fairness, and QoS targets simultaneously.

Concept diagram

CONTROLLER SCHEDULING LOOP

request queues -> row-policy + priority -> command issue -> bank state update

Metric graph

QUEUE PRESSURE MIX

row-hit preference bias ██████
aging/fairness pressure █████
QoS override cost       ███

Reports and artifacts

• scheduler policy comparison

• queue age distribution

• starvation/fairness incident report

• QoS latency percentile dashboard

Mini case study

FR-FCFS tuning improved bulk throughput but starved latency-critical traffic until age caps and class quotas were added.

Debug branches

• Measure queue age tails by traffic class

• Separate row-hit gains from fairness regressions

• Stress policy under mixed burst and random streams

Senior review question

Ask: which latency, bandwidth, and reliability evidence proves this DRAM topic is closed under real traffic?

Key takeaways

• Always tie controller and PHY counter shifts to application latency and throughput outcomes.

• Lock firmware timing profile, thermal condition, and DIMM state before comparing DRAM captures.

Common pitfalls

• Chasing peak bandwidth while ignoring p99 latency and fairness tails.

• Changing timing guardbands without separating SI noise from scheduling issues.

• Declaring closure without reliability gates, fault injection, and regression replay.

Principal DRAM review addendum

QoS Classes, Priority Arbitration, and Starvation Boundaries should be read as an end-to-end memory behavior, not as a single block definition. A production DRAM subsystem reflects interactions between array physics, command legality, scheduler policy, PHY margin, and reliability controls before software experiences final latency or bandwidth.

QoS-aware arbitration overlays policy on top of raw efficiency scheduling so critical clients (for example CPU demand fetches, display, or real-time accelerators) get bounded service even when background traffic is heavy. The controller typically uses weighted priority, aging, credit/token buckets, or deadline-aware boosts to pick among ready requests. Pure fixed priority can satisfy critical latency but often starves low-priority flows; pure fairness can miss hard deadlines. Practical designs combine tiers: first enforce hard constraints (deadline/critical window), then apply weighted fairness among remaining contenders, with aging to guarantee eventual service. Arbitration decisions must be synchronized with read/write batching, bus turnaround penalties, and bank availability, otherwise QoS policy can look correct at request level yet fail at command-level execution. End-to-end QoS therefore requires both scheduler logic and upstream traffic shaping: if NoC or cache eviction policy injects pathological bursts, controller-only fixes may be insufficient. Robust implementations validate SLA behavior using adversarial traffic mixes and explicitly monitor tail latency excursions, not just average service rate. DRAM inefficiency is multiplicative: one extra ACTIVATE, one unnecessary turnaround, one weak lane margin, or one refresh collision repeated across billions of accesses can dominate product tail latency and power.

Use Per-class latency SLA compliance (real-time, interactive, best-effort) and fairness index under stress traffic. as the opening signal, not the conclusion. A metric move only becomes actionable when paired with workload context, command traces, training telemetry, and evidence artifacts such as QoS compliance dashboard with per-class SLA miss counters, arbitration decision logs, and starvation watchdog events..

Memory-controller quality is measured by throughput and tail predictability under mixed traffic, not average bandwidth alone. Senior review quality comes from proving a complete chain: request pattern -> memory-state transition -> bottleneck mechanism -> smallest owner fix -> regression-safe validation.

Review discipline should enforce a single causal chain: traffic pattern -> command-level behavior -> array/PHY effect -> measured product impact. That chain prevents tuning folklore from replacing evidence.