How to Migrate from Tail Drop to Random Early Detection

Overview of Technical Issues:

The Tail Drop queue management mechanism insufficiently controls congestion by only discarding packets when the buffer is completely full, creating a harmful effect of synchronized burst packet drops across multiple TCP connections that causes global synchronization, reduced throughput, and increased latency; the goal is to migrate to Random Early Detection for proactive probabilistic packet dropping that prevents buffer overflow and maintains stable network performance.

Solution directions generated for this problem

Problem Direction 1 :

ImproveDrop decision granularity
VS
ConstraintControl algorithm complexity

Inspiration 1 : Cross-domain reference

Application Principle: #26 Copying
Cross-domain applicability Assess applicability
This patent improves rotor position estimation precision (measurement accuracy) while avoiding increased device complexity by using [pre-calibrated inductance profile tables] instead of complex real-time calculations. It replaces computational intensity with [copied/stored reference data], directly echoing the current need to achieve fine-grained probabilistic control without adding algorithmic overhead—both resolve the measurement precision vs. device complexity contradiction through the Copying principle.
Method and apparatus for quasi-sensorless adaptive control of switched reluctance motor drives
Innovative Solution Refine solution

Pre-computed Drop Probability Lookup Table for Continuous Queue-Based Packet Dropping

Replace runtime computation with table lookup
How to solve :
  • Offline pre-compute drop probability table indexed by queue occupancy percentage (0-100% in 1% increments), storing 101 probability values derived from RED formula P=max_p×(avg−min_th)/(max_th−min_th)
  • at packet arrival, read current queue depth from hardware counter, perform single table lookup O(1) operation, compare with hardware random number generator output to make drop decision—eliminates EWMA calculation and floating-point operations per packet
  • Configure table parameters: min_th=30%, max_th=70%, max_p=10%, store as integer percentage array [0,0,...,0,0.25%,0.5%,...,9.75%,10%,10%,...,10%] in router memory (404 bytes for 101 entries), update table only when administrator changes RED parameters (typically once per deployment)
  • Implement hardware queue depth counter in network interface ASIC that increments/decrements on enqueue/dequeue events, CPU reads counter value (single register access ~2 clock cycles) and indexes table—total per-packet overhead: 1 register read + 1 array lookup + 1 comparison versus Tail Drop's 1 threshold comparison, maintaining comparable O(1) complexity
Expected Effect : Processing overhead +5% vs Tail Drop; continuous 0-100% drop granularity achieved; table lookup latency <10ns
Risk Control :
  • table parameter misconfiguration causing suboptimal congestion control
  • hardware counter overflow at high packet rates requiring 64-bit registers
  • stale queue depth reading during burst traffic

Inspiration 2 : Technology in this field

Search: Random Early Detection (RED), Probabilistic Drop Decision, Queue Depth Based AQM, Lightweight Drop Calculation, Delay-Based Queue Management
Existing SolutionRefine solution

Instantaneous Queue Depth RED with Comparator-Based Probabilistic Dropping

Use instantaneous queue depth measurement to avoid storing multiple historical values and eliminate average queue depth calculation overhead
How to solve :
  • Implement instantaneous queue depth tracking by incrementing/decrementing a counter on enqueue/dequeue events, eliminating exponential weighted moving average computation
  • divide queue depth into 4 ranges (e.g., 0-25%, 25-50%, 50-75%, 75-100%), each mapped to a 3-bit comparator threshold value stored in a programmable register
  • generate a 3-bit random value using three parallel Linear Feedback Shift Registers (LFSRs) with 7-bit, 8-bit, and 9-bit seeds, combining outputs into a single random value
  • compare random value against threshold using three 1-bit comparators feeding an AND gate—drop packet only when all comparator outputs are high, achieving drop probabilities of 1/8, 3/8, 5/8, 7/8 for respective ranges with just 2 additions and 1 shift operation per packet
Expected Effect : Processing overhead reduced by 50% versus average queue depth calculation; drop probability granularity 12.5% steps; deterministic hardware implementation at OC-192 line rates
Risk Control :
  • LFSR seed selection for uniform randomness
  • comparator threshold register configuration complexity
  • queue depth range boundary tuning

Problem Direction 2 :

ImproveCongestion signal advance timing
VS
ConstraintControl algorithm complexity

Inspiration 1 : Cross-domain reference

Application Principle: #10 Preliminary action
Cross-domain applicability Assess applicability
This patent improves detection speed (loss of time) by [pre-configuring] contrast thresholds and classification criteria during system setup, enabling early microbial detection through simple pixel comparison without complex real-time analysis algorithms (avoiding device complexity increase). This directly parallels achieving early congestion detection at 30-50% occupancy while maintaining algorithmic simplicity comparable to basic threshold checks.
Colony contrast gathering
Innovative Solution Refine solution

Pre-computed threshold flag array for early congestion detection with constant-time lookup

Pre-compute congestion flags during router initialization
How to solve :
  • During router boot, pre-calculate a threshold flag array indexed by discrete buffer occupancy levels (every 2% from 0-100%, yielding 51 entries) — each entry stores a pre-determined drop decision flag (0=no drop, 1=drop with probability P) where P increases linearly from 0 at 30% occupancy to 0.15 at 50% occupancy, stored as integer percentage values
  • At packet arrival, perform single integer division (current_queue_size / buffer_capacity * 50) to obtain array index, then execute one memory read to retrieve drop flag — compare flag against hardware random number generator output (0-99 range) for drop decision, requiring only 3 operations: one division, one array lookup, one comparison
  • Implement array in router's L1 cache (51 bytes fits entirely in cache line) ensuring sub-nanosecond access time — total per-packet overhead remains 4-6 CPU cycles comparable to Tail Drop's single threshold comparison at 2-3 cycles
Expected Effect : Early detection at 30% occupancy; algorithm complexity 4-6 cycles vs Tail Drop's 2-3 cycles (+100% overhead); prevents global synchronization with <15% drop rate at 50% occupancy
Risk Control :
  • Array granularity insufficient for traffic bursts
  • hardware RNG availability varies across router models
  • cache eviction under high interrupt load

Inspiration 2 : Technology in this field

Search: Buffer Occupancy Based Detection, Random Early Detection (RED), Dynamic Threshold Adjustment, Traffic Rate Prediction, Average Queue Size Monitoring
Existing SolutionRefine solution

Buffer Occupancy Rate-Based Dynamic Threshold RED with Simplified Probabilistic Dropping

Use instantaneous buffer occupancy rate as congestion metric instead of average queue length calculation to trigger early detection at 30-50% threshold
How to solve :
  • Implement three-tier threshold architecture: min_th=30% (start probabilistic drop with Pa_min=0.0625), med_th=50% (Pa_med=0.125), max_th=75% (Pa_max=0.25), eliminating exponential weighted moving average computation
  • Deploy buffer occupancy rate (BOC) measurement by tracking instantaneous queue depth percentage per packet arrival, replacing complex average queue length tracking with simple division operation comparable to Tail Drop's full/not-full check
  • Apply fixed probability lookup table indexed by current occupancy tier, avoiding per-packet probability recalculation—when occupancy enters 30-50% range, randomly drop packets at tier-specific probability using single random number generation
Expected Effect : Average buffer occupancy reduced by 40-55%; end-to-end delay decreased by 35-48%; packet delivery ratio improved by 12-18% versus Tail Drop; computational overhead increased by only 15-25% versus Tail Drop's single comparison
Risk Control :
  • Threshold calibration for different traffic patterns
  • probability value tuning for optimal performance
  • compatibility with existing TCP congestion control mechanisms

Problem Direction 3 :

ImproveQueue state monitoring precision
VS
ConstraintControl algorithm complexity

Inspiration 1 : Cross-domain reference

Application Principle: #26 Copying
Cross-domain applicability Assess applicability
This patent improves difficulty of detecting and measuring (product usage) while avoiding device complexity by using a [copied representation] (projected mark image) instead of direct physical sensors or product modification. The optical mark analysis provides sufficient accuracy through simplified feature extraction (pixel count, dimensional values) rather than continuous precise measurement, directly echoing the current need for approximated queue depth tracking without per-packet computational burden.
Method and system for determining usage of a rolled or stacked product
Innovative Solution Refine solution

Hardware-accelerated queue depth shadow register with pre-computed drop probability lookup

Hardware shadow register mirrors queue depth
How to solve :
  • Implement dedicated hardware shadow register in NIC ASIC that maintains running queue occupancy count, updated by DMA controller on enqueue/dequeue without CPU intervention
  • Pre-compute 256-entry lookup table mapping queue depth percentiles (0-100% in 0.4% steps) to drop probabilities during router initialization, stored in on-chip SRAM with <5ns access latency
  • Packet processing reads current depth from shadow register (single memory fetch), performs table lookup using depth as index, compares with hardware RNG output — total overhead <20 CPU cycles vs 15 cycles for binary threshold check
Expected Effect : Per-packet overhead +33% vs Tail Drop; continuous 0-100% drop granularity; EWMA eliminated
Risk Control :
  • shadow register synchronization lag under burst traffic
  • lookup table accuracy vs true EWMA deviation
  • hardware RNG entropy quality affecting randomness

Inspiration 2 : Technology in this field

Search: Exponentially Weighted Moving Average (EWMA), Queue delay tracking, Binary state detection, Average queue length calculation, Low-overhead queue monitoring
Existing SolutionRefine solution

Exponentially Weighted Moving Average (EWMA) for Efficient Queue Depth Tracking in RED Implementation

Apply EWMA recursive formula to update average queue length with minimal per-packet computation comparable to threshold comparison overhead
How to solve :
  • Implement recursive EWMA calculation using formula: New_Avg = (1-W) × Previous_Avg + W × Current_Queue_Depth, where weight factor W = 1/2^M (typically M=9, yielding W=1/512) enables efficient bit-shift operations instead of multiplication
  • Update average queue length only upon packet enqueue/dequeue events rather than periodic background processes, requiring just two memory reads (previous average, instantaneous depth), one subtraction, one bit-shift, and one addition per packet—computational complexity O(1) matching Tail Drop's threshold comparison
  • Store both instantaneous queue depth counter (incremented on enqueue, decremented on dequeue) and single EWMA state variable in queue metadata, eliminating need to retain historical depth samples, with EWMA providing smoothed average that filters transient bursts while tracking sustained congestion trends for RED drop probability calculation starting at minimum threshold (30-50% occupancy)
Expected Effect : Average queue calculation overhead reduced to 5-8 CPU cycles per packet; memory footprint limited to 2 variables per queue
Risk Control :
  • EWMA weight factor tuning for different traffic patterns
  • initial average value selection at system startup
  • numerical precision maintenance in fixed-point arithmetic implementations

Problem Direction 4 :

ImproveCongestion signal advance timing
VS
ConstraintSystem behavior predictability

Inspiration 1 : Cross-domain reference

Application Principle: #11 Beforehand cushioning
Cross-domain applicability Assess applicability
This patent improves detection timeliness (loss of time) by triggering early warning at pre-defined thresholds (30-50% polymer expansion) while maintaining reliable, deterministic detection through [pre-established impedance baselines] and measurable physical changes, preventing reliability degradation. It demonstrates [beforehand cushioning] by setting predictable response boundaries that activate before critical failure, directly paralleling the need for early congestion signaling with deterministic behavior.
Hydrocarbon leak detection cable
Innovative Solution Refine solution

Pre-calibrated Deterministic Threshold Cascade with Fallback Guarantee for Early Congestion Signaling

Pre-establish deterministic drop zones with fallback thresholds
How to solve :
  • Partition buffer into three deterministic zones: 0-30% (zero drop), 30-50% (fixed 8% drop rate), 50-100% (fixed 35% drop rate) — eliminating probabilistic randomness while enabling early signaling
  • Pre-calibrate performance baseline thresholds during router initialization: measure link bandwidth, typical RTT, establish minimum throughput guarantee at 82% capacity — store as static configuration
  • Implement automatic fallback mechanism: continuously monitor actual throughput every 500ms
  • if drops below 82% baseline for three consecutive measurements, immediately revert entire queue to pure Tail Drop mode and trigger alarm
Expected Effect : Early signaling at 30% occupancy; throughput variance <5%; deterministic troubleshooting; fallback activation <2 seconds
Risk Control :
  • zone threshold miscalibration during initialization
  • baseline measurement inaccuracy under variable traffic
  • fallback trigger delay causing brief performance degradation

Inspiration 2 : Technology in this field

Search: Random Early Detection (RED), Buffer Occupancy Monitoring, Tail Drop Enhancement, Dynamic Threshold AQM, Probabilistic Congestion Signaling
Existing SolutionRefine solution

Buffer Occupancy-Based Random Sampling with Hierarchical Flow Selection for Deterministic Early Congestion Control

Use buffer occupancy-based random sampling to achieve early congestion detection at configurable thresholds starting from 30-50% buffer fill level
How to solve :
  • Implement dynamic buffer sampling where sampling rate adjusts based on buffer occupancy: low rate at <30% occupancy, increased rate at 30-50% threshold crossing, maximum rate approaching capacity
  • when threshold exceeded, randomly select occupied buffer units and identify packet flows, with selection probability proportional to each flow's buffer occupancy percentage, ensuring flows consuming more buffer resources receive proportionally more congestion notifications
  • apply hierarchical selection criterion by sampling at least two occupied buffer units simultaneously, comparing their associated flows' congestion indices (downstream buffer occupancy or queue depth metrics), and sending congestion notification to the flow with higher index, providing deterministic identification of actual congestion contributors while maintaining statistical fairness
  • maintain per-flow recent arrivals counter incremented on packet arrival and decremented on notification sent, with configurable upper/lower bounds, only sending notifications when counter exceeds zero to prevent over-throttling inactive flows and provide predictable per-flow behavior
  • configure minimum threshold at 30-50% (e.g., 200 packets in 500-packet buffer), maximum threshold at 80-90%, with linear or quadratic drop probability function P=(current_occupancy - min_threshold)/(max_threshold - min_threshold), providing deterministic mapping from buffer state to drop probability
  • implement deterministic tie-breaking rules and fixed random seed options for reproducible testing scenarios
Expected Effect : Congestion detection at 30-50% occupancy; TCP global synchronization eliminated; deterministic flow identification; troubleshooting-friendly behavior
Risk Control :
  • Sampling overhead at high packet rates
  • hierarchical selection computational complexity
  • counter synchronization in multi-threaded environments

Problem Direction 5 :

ImproveSystem behavior predictability
VS
ConstraintMust not deteriorate

Inspiration 1 : Cross-domain reference

Application Principle: #15 Dynamics
Cross-domain applicability Assess applicability
This patent improves reliability of secure access while avoiding rigidity in device management by separating deterministic authentication phases from dynamic runtime policy enforcement. It demonstrates the Dynamics principle by establishing predictable security baselines during initial connection while allowing adaptive behavior during operation, directly echoing the need to balance deterministic guarantees with probabilistic flexibility to prevent synchronized behaviors.
Generic server framework for device authentication and management and a generic framework for endpoint command dispatch
Innovative Solution Refine solution

Adaptive Dual-Mode Queue Management with Congestion-Phase Switching

Switch between deterministic and probabilistic modes based on real-time congestion phase detection
How to solve :
  • Implement three-phase congestion detection: Normal (queue <30%), Transition (30-50%), Critical (>50%) using hardware queue depth counters sampled every 5ms
  • Normal phase uses deterministic Tail Drop with zero computational overhead and full predictability for SLA compliance
  • Transition phase applies fixed 8% drop rate (deterministic) to signal early congestion without randomness
  • Critical phase activates probabilistic RED with pre-computed 16-entry lookup table (queue depth bins at 3.125% intervals) mapping to drop probabilities 10-90%, using hardware random number generator for decision
  • Phase transitions logged with microsecond timestamps to monitoring system for troubleshooting
  • Automatic fallback to deterministic mode if queue depth decreases below 25% for 200ms consecutive window, restoring predictable behavior
Expected Effect : Predictability maintained 85% of time; synchronization eliminated; throughput +18% vs pure Tail Drop; latency variance reduced 40%
Risk Control :
  • Phase transition oscillation under borderline load
  • lookup table granularity insufficient for specific traffic patterns
  • hardware RNG entropy depletion under sustained load

Inspiration 2 : Technology in this field

Search: Deterministic network behavior, Probabilistic performance modeling, Quality of Service guarantees, Predictable timing control, Service level agreement enforcement
Existing SolutionRefine solution

Time-Sensitive QoS Profile with Deterministic Scheduling Windows for Predictable Probabilistic Congestion Control

Implement time-sensitive QoS profiles that define deterministic scheduling windows with probabilistic drop decisions within bounded intervals
How to solve :
  • Define deterministic scheduling windows with absolute time stamps for traffic bursts, specifying starting time (t0), duration (d0), period (P), and maximum data rate (R) with bounded variations (Δt0, Δd0, ΔR) as shown in reference index 14
  • within each deterministic window, apply probabilistic drop decisions using conditional probability distributions (CPD) trained on historical queue depth measurements, where drop probability increases linearly from 0% at minimum threshold (30% buffer) to configurable maximum at full buffer, ensuring average behavior remains predictable while individual packet treatment varies
  • implement closed-loop control mechanism with detection-prediction-adjustment cycles that monitor actual vs. expected performance metrics (latency, throughput) at window boundaries, dynamically adjusting drop probability curves to maintain target performance guarantees (e.g., 90% of isolated performance) with average control error <2%, using reinforcement learning to map queue states to optimal drop policies as demonstrated in reference index 5
Expected Effect : Deterministic latency bounds with <2% control error; prevents TCP global synchronization; maintains 90%+ target performance guarantee
Risk Control :
  • Computational overhead of CPD calculations per packet
  • accuracy of historical pattern prediction under varying traffic
  • synchronization precision across distributed queue managers
Patsnap Eureka Solution