1,000,000 transactions per second, measured, not marketed
MatchCore's cluster throughput scales close to linearly with node count. Here's the methodology behind the headline number, and the latency you should expect at every scale.
Throughput scales with cluster size
Percentiles, per segment, under sustained load
| Percentile | Wall-to-wall latency | Load condition |
|---|---|---|
| p50 | <12µs | 150,000 orders/sec sustained |
| p95 | <28µs | 150,000 orders/sec sustained |
| p99 | <50µs | 150,000 orders/sec sustained |
| p99.9 | <140µs | 150,000 orders/sec sustained, including GC-quiesced runtime |
How we measure it
Reference hardware
Bare-metal nodes, dedicated NICs with kernel-bypass networking, NVMe-backed write-ahead log storage, one matching shard pinned per physical core.
Workload profile
Synthetic order flow modelled on lit-market equities activity: a realistic mix of new orders, cancels, and modifies at production-representative price-level depth.
What "wall-to-wall" means
Time from inbound gateway receipt to outbound execution report, including protocol parsing and market-data publication — not just core matching time.
Your workload may differ
Throughput and latency depend heavily on instrument count, order-to-trade ratio, and message size. We benchmark against your own flow profile on request.