Random Early Detection for Multi-Tenant Cloud Networks

Overview of Technical Issues:

In multi-tenant cloud networks, the packet dropping controller applying traditional Random Early Detection algorithm insufficiently differentiates between tenant traffic flows, causing aggressive tenants to fill shared queues and trigger indiscriminate packet drops that penalize all tenants equally; the goal is to achieve fair bandwidth allocation and proportional packet drop rates across tenants while maintaining congestion control effectiveness.

Solution directions generated for this problem

Problem Direction 1 :

ImproveTenant traffic differentiation granularity
VS
ConstraintController computational complexity

Inspiration 1 : Cross-domain reference

Application Principle: #1 Segmentation
Cross-domain applicability Assess applicability
This patent improves calculation precision (vector position determination) while preventing device complexity deterioration by using segmented coordinate system decomposition to avoid expensive trigonometric operations, directly paralleling the current contradiction of achieving precise per-tenant measurement while maintaining O(1) controller complexity through segmentation.
A SVPWM modulation method based on a five-segment five-level converter
Innovative Solution Refine solution

Hierarchical hash-indexed tenant metering with spatial decomposition

Decompose tenant space into hash buckets for parallel constant-time lookup
How to solve :
  • Partition N tenants into M fixed hash buckets (M=16–64) using tenant_ID modulo M, each bucket maintains aggregate queue occupancy and arrival rate counters updated in hardware
  • Controller performs single hash lookup O(1) to retrieve bucket statistics, applies drop probability to packet based on bucket-level metrics rather than individual tenant state
  • Implement hardware counter arrays in programmable switch registers (P4) or FPGA block RAM, with atomic increment operations triggered by packet metadata at line rate without CPU involvement
Expected Effect : Complexity O(1) constant; latency <500ns; fairness error <8%
Risk Control :
  • hash collision causing bucket imbalance
  • counter overflow at high traffic rates
  • bucket size miscalibration

Inspiration 2 : Technology in this field

Search: O(1) constant-time complexity, packet scheduling algorithms, load-balanced switch architecture, constant-size state cache
Existing SolutionRefine solution

Periodic State Update Mechanism for Constant-Time Per-Tenant Traffic Monitoring

Apply periodic state update mechanism inspired by TConstFormer architecture to maintain per-tenant traffic statistics in constant-size memory
How to solve :
  • Implement k-periodic update mechanism where k=256-512 packets: perform O(1) constant-time incremental counter updates (packet count, byte accumulation) for k-1 consecutive packets using tenant-ID indexed hash table with fixed bucket size
  • execute single O(N) global synchronization on k-th packet to calculate per-tenant queue occupancy ratios and arrival rates, amortizing complexity to O(1) per packet
  • deploy dual-buffer state structure with active monitoring buffer for real-time O(1) updates and shadow calculation buffer for periodic aggregation, ensuring microsecond-scale packet processing latency remains unaffected during synchronization cycles
  • utilize fixed-size circular hash table with 1024-2048 buckets using tenant-ID modulo mapping, collision resolution via linear probing limited to 4 slots maximum, guaranteeing worst-case O(1) lookup with ≤5 memory accesses
Expected Effect : Amortized O(1) complexity per packet; 99.6% operations pure constant-time; memory footprint 8-16KB fixed
Risk Control :
  • Hash collision handling under tenant churn
  • synchronization jitter impact on fairness accuracy
  • memory access latency variance

Problem Direction 2 :

ImprovePacket drop fairness across tenants
VS
ConstraintSystem response latency

Inspiration 1 : Cross-domain reference

Application Principle: #10 Preliminary action
Cross-domain applicability Assess applicability
This patent improves processing reliability (parallel filter application accuracy) while preventing loss of time (processing delays) by [pre-determining] filter applicability using reference pixels before actual filtering execution. This mirrors the current contradiction of achieving reliable proportional drop rates without microsecond-scale latency degradation through preliminary computation.
Image processing device and image processing method that horizontal filtering on pixel blocks
Innovative Solution Refine solution

Pre-computed tenant drop probability cache with asynchronous refresh for microsecond-scale fair congestion control

Cache tenant drop probabilities offline for instant lookup
How to solve :
  • Pre-calculate per-tenant drop probability tables every 5ms based on recent queue occupancy ratios (trailing 10ms window), storing N tenant entries in on-chip SRAM cache with 8-byte records (tenant_ID + drop_prob)
  • Upon packet arrival, perform single O(1) hash lookup using tenant_ID from packet header metadata to retrieve cached probability, execute drop decision via hardware random number generator comparison within 2 microseconds
  • Run asynchronous background thread on separate CPU core to refresh cache entries every 5ms by computing weighted moving average of tenant byte contributions (α=0.7 for recent samples), updating drop probabilities proportional to each tenant's queue share deviation from fair allocation
Expected Effect : Drop decision latency <3μs, fairness coefficient >0.95, cache hit rate >99.8%
Risk Control :
  • cache coherency during refresh cycles
  • hash collision with high tenant count
  • stale probability during traffic bursts

Inspiration 2 : Technology in this field

Search: Latency-based congestion control, Probabilistic packet drop, Multi-tenant fairness, Proportional rate control, Microsecond-scale response
Existing SolutionRefine solution

Priority-Weighted WRED for Per-Tenant Proportional Packet Dropping

Extend WRED with tenant priority classification to achieve proportional drops without separate queues
How to solve :
  • Assign each tenant a priority weight σᵢ=1/priceᵢ based on service tier
  • implement service priority logic within WRED congestion window (MINTHRES to MAXTHRES) where drop probability per tenant pd,i = (σᵢ/Σσⱼ)·pd_overall, ensuring lower-priority tenants drop first
  • classify packets via VLAN IDs or tenant identifiers in packet headers at line-rate using hardware classification circuits, maintaining O(1) per-packet processing
Expected Effect : Proportional drop rates across tenants; congestion response latency <10μs maintained; link utilization >95%
Risk Control :
  • Priority assignment policy fairness
  • hardware classification circuit scalability to thousands of tenants
  • interaction with TCP congestion control algorithms

Problem Direction 3 :

ImproveBandwidth allocation proportionality
VS
ConstraintController computational complexity

Inspiration 1 : Cross-domain reference

Application Principle: #35 Parameter changes
Cross-domain applicability Assess applicability
This patent improves trajectory planning productivity (real-time performance) while preventing device complexity deterioration by using [segmented planning] that computes parameters only at critical transition points rather than continuously. This directly echoes the current contradiction of achieving fair bandwidth allocation (productivity) while keeping controller complexity low by adjusting calculation [granularity] from per-packet to per-epoch intervals.
A trajectory planning method and device
Innovative Solution Refine solution

Epoch-based hierarchical tenant bandwidth allocation with adaptive refresh intervals

Shift from per-packet to epoch-based allocation
How to solve :
  • Divide time into adaptive epochs (baseline 10ms, shortened to 2ms under congestion detected by aggregate queue depth >70%)
  • compute per-tenant bandwidth shares only at epoch boundaries using hardware byte counters, reducing calculation frequency by 1000–5000×
  • Implement two-tier tenant grouping: classify N tenants into log₂(N) priority bins (4–6 bins typical) based on SLA tier and recent traffic volume
  • allocate bandwidth among bins first (O(log N)), then equally within bins, avoiding per-tenant iteration
  • Deploy hardware token bucket enforcers in programmable switches (P4) or SmartNICs: controller updates per-bin rate limits once per epoch via control plane API
  • data plane enforces limits at line rate using pre-loaded token rates, decoupling fairness computation from packet processing
Expected Effect : Controller complexity O(log N); fairness error <8%; latency <50μs
Risk Control :
  • epoch duration miscalibration under traffic bursts
  • bin boundary oscillation causing tenant migration overhead
  • hardware counter overflow at high packet rates

Inspiration 2 : Technology in this field

Search: Max-Min Fair Allocation, Proportional Fairness, Stateless Queue Management, Tenant Bandwidth Cap, Dynamic Pricing Control
Existing SolutionRefine solution

Approximate Fair Dropping with Stateless Per-Tenant Rate Limiting

AFD uses stateless per-tenant flow tracking to achieve fair bandwidth allocation without per-flow queuing complexity
How to solve :
  • Implement stateless matched-drop strategy where incoming packets are probabilistically compared against randomly sampled buffered packets from the same tenant flow
  • if match occurs, drop both packets with probability proportional to tenant's queue occupancy ratio, achieving approximate weighted max-min fairness
  • Maintain per-tenant byte counters updated on packet arrival/departure using hash-based flow identification (5-tuple hashing), enabling O(1) lookup complexity
  • no per-flow state storage required
  • Apply dynamic drop probability calculation as P_drop(tenant_i) = min(1, (bytes_i / total_bytes) × (queue_length / threshold)), where threshold set at 50-70% of buffer capacity to trigger early congestion response while maintaining link utilization above 95%
Expected Effect : Fair bandwidth allocation within 5% of ideal max-min fairness; O(1) per-packet processing overhead
Risk Control :
  • Hash collision handling for flow identification accuracy
  • tuning drop probability parameters for diverse traffic patterns
  • backward compatibility with existing TCP congestion control

Problem Direction 4 :

ImproveTenant traffic differentiation granularity
VS
ConstraintSystem response latency

Inspiration 1 : Cross-domain reference

Application Principle: #10 Preliminary action
Cross-domain applicability Assess applicability
This patent improves prediction accuracy and response speed by [pre-computing] personalized language models and [pre-storing] candidate hypotheses in optimized data structures (range trie), enabling real-time lookup without sacrificing interaction latency. It demonstrates how preliminary action resolves the contradiction between improving measurement precision (accurate auto-completion) while preventing loss of time (real-time response).
Contextual auto-completion for assistant systems
Innovative Solution Refine solution

Pre-computed tenant drop probability lookup table with asynchronous refresh for microsecond-scale congestion control

Pre-compute tenant metrics offline for real-time lookup
How to solve :
  • Deploy dual-buffer tenant probability table in shared memory — buffer A serves packet drop decisions via O(1) hash lookup while buffer B updates asynchronously every 100μs based on hardware counter snapshots
  • Implement hardware-accelerated tenant counters using NIC SR-IOV virtual function registers or programmable switch stateful ALUs to track per-tenant byte counts and packet arrivals at line rate without CPU involvement
  • Execute background refresh thread on dedicated CPU core that reads hardware counters every 100μs, recalculates N tenant drop probabilities using weighted RED formula, and atomically swaps buffer pointers — packet processing path only performs tenant_id hash lookup (≤50ns) plus random number comparison (≤30ns)
Expected Effect : Per-packet latency ≤200ns; fairness update cycle 100μs; supports 10K tenants
Risk Control :
  • hash collision rate exceeds 1%
  • buffer swap atomicity violation
  • hardware counter overflow

Inspiration 2 : Technology in this field

Search: Per-tenant queue management, Tenant resource usage measurement, Queue occupancy monitoring, Low-latency admission control, Arrival rate modeling
Existing SolutionRefine solution

Hierarchical Snapshot-Based Per-Tenant Queue Occupancy Tracking with Hash-Indexed Measurement

Implement time-windowed snapshot architecture that segments queue departure traffic into fixed intervals and maintains per-tenant flow counters using hash-indexed data structures
How to solve :
  • Deploy Count-Min Sketch (CMS) with 4-row hash tables (24-32 counters per snapshot, <2KB memory) to record per-tenant packet counts in rotating time windows T (100μs-1ms intervals)
  • maintain h snapshots (h ≥ Q/(R×T), where Q=buffer size, R=line rate) in round-robin fashion with concurrent Write/Read/Clean phases
  • calculate per-tenant queue occupancy by aggregating snapshots covering arrival-to-departure interval [ai, di] using hash lookups on tenant flow identifiers
  • implement snapshot recycling without controller intervention by incrementing version counters and cleaning expired snapshots via cyclic pointer advancement
  • parallelize snapshot operations across pipeline stages by stacking multiple hash arrays in same stage to meet <10μs processing deadlines
Expected Effect : >90% precision/recall in culprit flow detection; <2KB memory per port; sub-millisecond measurement granularity
Risk Control :
  • Hash collision management under high tenant counts
  • snapshot synchronization timing accuracy
  • memory bandwidth constraints in multi-pipeline architectures
Patsnap Eureka Solution