8. Performance
8.1 Throughput Targets
The repository contains benchmark targets and benchmark programs, but the source snapshot does not establish the following numbers as measured production performance. They must be reproduced on a published commit, hardware profile, network topology, transaction mix, and validator count.
| Metric | Current status |
|---|---|
| Per-shard throughput | Target only; no normative measured number |
| Consensus latency | Configuration- and network-dependent target |
| Finality time | Depends on checkpoint interval, validator participation, and network conditions |
| Aggregate throughput | A scaling hypothesis; not a guaranteed linear function of shard count |
Cross-shard execution, signature verification, storage I/O, network propagation, and checkpoint formation can all reduce throughput. A valid benchmark report must include failed transactions, conflict rate, cross-shard ratio, resource usage, and finality latency; a synthetic loop benchmark is not sufficient evidence for a chain-wide TPS claim.
8.2 Signature Overhead
Signature overhead must be calculated from the concrete ML-DSA-65 implementation and workload used by the node, together with ML-KEM operations and proof/aggregation helpers. A valid report must publish signature sizes, verification time, bandwidth, storage retention, cache effects, and whether raw signatures or commitments are propagated.
8.3 Where the Throughput Comes From
Per-shard throughput is not achieved by a single trick but by stacking several parallelism mechanisms, each described in detail in later sections:
- Inter-shard parallelism (Dynamic Sharding section): independent shards execute simultaneously; aggregate throughput scales with active shard count.
- Intra-shard execution parallelism (Virtual Machine section): within a shard, the transaction dependency graph, MVCC snapshot isolation, and speculative execution let non-conflicting transactions run concurrently across cores instead of sequentially.
- Signature commitments (Cryptography section): aggregation and Merkle-commitment helpers can reduce representation size, but the raw signature verification and archival costs remain part of the benchmark.
- Crypto acceleration: SIMD, precomputation, pooling and batch-verification helpers are available; their speedup must be measured for the concrete profile.
The realistic figure to reason about is therefore per-shard throughput under a given cross-shard ratio; the aggregate number follows from how many shards are active and how localised the workload is.