Mempool Architecture

The mempool manages pending transactions from admission through scoring to block production draining.

Pipeline

Network (gossipsub /savitri/tx/1)
    │
    ▼
Prevalidation (stateless)
    │ signature format, size limits
    ▼
Admission Control
    │ quota check, per-sender cap
    ▼
Stateful Validation
    │ nonce, balance, fee limits
    ▼
SIMD Batch Scoring
    │ fee * 0.7 + class * 0.3
    ▼
Score Cache (LRU)
    │
    ▼
Pending Pool
    │
    ▼
drain_for_block_production()
    │ top-N by score, final_validation
    ▼
Block Proposer

Transaction Classes

Class	Priority	Use Case
FederatedUpdate	1.0 (highest)	FL model updates
Financial	0.8	Token transfers
Governance	0.7	Votes, proposals
IoTData	0.5	Sensor data

Admission Quotas

Class	Max Pending	Per-Sender Cap
Financial	50,000	512
IoT	100,000	512
Global	100,000	512

The drain_fair_batch() function calls record_removal() to properly decrement class counts when transactions are drained (fixed in Round 6).

SIMD Scoring

Score Formula

score = fee_normalized * fee_weight + class_priority * class_weight

Default: fee_weight = 0.7, class_weight = 0.3.

Architecture-Specific Implementation

Architecture	Intrinsics	Lanes	Min Batch
x86_64 (AVX2+FMA)	_mm256_set1_pd, _mm256_fmadd_pd	4 doubles	32
ARM (NEON)	vdupq_n_f64, vfmaq_f64	2 doubles	32
Fallback	Scalar loop	1	Always

For batches < 32 transactions, scalar computation is used because SIMD setup overhead exceeds the benefit.

Adaptive Weights (Optional)

With adaptive_weights feature flag, weights adjust based on:

• Fee distribution: If fees are clustered high, increase fee_weight
• Class diversity: If many different classes, increase class_weight
• Historical throughput: Adjust based on recent blocks

Parameters:

base_fee_weight = 0.7
base_class_weight = 0.3
adaptation_rate = 0.1 (smoothing)
fee_threshold_high = 2,000,000,000
fee_threshold_low = 500,000,000
class_diversity_threshold = 0.5

Score Cache

LRU cache avoids recomputing scores:

Parameter	Development	Production
Cache size	100 entries	10,000 entries
TTL	60 seconds	300 seconds

Nonce Management

Standard Flow

1. TX arrives with nonce N
2. Check: N >= account.storage_nonce
3. Track in pending_nonces HashMap during draining
4. Commit nonce after block finalization

Gap Tolerance

Nonce gap up to 5000 is accepted for burst recovery after prolonged stalls (increased from 1000 in Round 11).

Cold-Start Fix

When storage_nonce=0 AND pending_nonces=0, the lowest available nonce is accepted as the starting point. This handles the case where TXs were drained but the block was never committed by a previous proposer.

Nonce Reset

TX generator resets local nonce when local_nonce > storage_nonce + 200 (MAX_NONCE_AHEAD). Previously local nonces could drift to 2000+ while storage was at 84, causing all TXs to be rejected.

Replay Prevention

Transaction hashes are tracked to prevent replay attacks. Once a TX hash is seen, duplicate submissions are rejected.

Fee Validation

FeeLimits {
    min_fee: 100_000_000_000_000,       // 0.0001 SAVT
    max_fee: 1_000_000_000_000_000_000, // 1.0 SAVT
}

Feature Flags

Flag	Description
simd	SIMD-optimized scoring (default)
cache	Score cache (default)
adaptive_weights	Dynamic weight adjustment