3.2 Throughput and latency

Our rules for numbers

  • Each number has a date and a source.
  • We do not mix live numbers and numbers from a model.
  • We remove artifacts that increase or decrease a number. We tell what we removed.
  • We give latency as p50 and p99. Half of the orders are faster than p50. One order in 100 is slower than p99.

Throughput

On 1 Nov 2026, a load test ran on an earlier chain without policing:

Item Value
Rate 1,035 commands each second, for 120 s
Commands executed with proof 145,235 of 145,235
LATE orders, full tables, relay overflows 0
Credits rejections, sent again by the clients 2,105

We did not do a load test on the policing chain.

Latency on the last chain (13 to 20 Nov 2026)

Leg p50 p99
CANON to executed 135.51 ms 256.81 ms
Executed to batch close 4.62 ms 20.69 ms
Batch close to FINAL 72.27 ms 133.31 ms
Full path (calculated, with the wait for the watermark) approximately 226 ms

This latency is too high, and we know the cause. On 3 Nov 2026, before policing, the full path was 22.5 ms at p50. On 6 Nov 2026, the stake rules started. With these rules, the host does not execute a batch before the previous batch has its certificate. Because of this, the batches do not overlap. Each batch waits for the FINAL of the previous batch (72 ms), and then for two evidence steps in the host (21 ms).

The data of all 9,053 minutes agrees:

Item, median of all minutes Value
Time between two batches 104.2 ms
Batch close to FINAL 72.4 ms
Host evidence steps for each batch 21.3 ms
Stake preparation for each batch 2.6 ms
CANON to executed (time that an order waits for its batch) 134.5 ms

Because of this, an order waits 134.5 ms for its batch. This wait is most of the 226 ms. The problem is in the host, not in the card.

What we will do. We will change the stake rules in the host. A batch will use the certificate of an older batch, not of the previous batch. Then the batches overlap again, as on 3 Nov 2026. We expect the full path to go back to approximately 22.5 ms, the value before the stake rules.

This change also changes the time when a stake penalty starts. It must have a complete design before we build it. We did not start the design.

Earlier measurements

Date Setup Full path, p50 / p99
28 Oct 2026 One order in each batch, 3 ms watermark 10.4 / 14.8 ms
3 Nov 2026 Batches in parallel, 10 ms watermark, no policing 22.5 / 38.1 ms
13 to 20 Nov 2026 Policing approximately 226 ms / not measured

Less than 1 ms: the design budget

The design target for the full path is less than 1 ms from the wire stamp to FINAL. The testnet does not meet it today. The table shows where the time goes now, and the design target for each part.

Part of the path Last chain, p50 Design target What must change
Wait for the watermark 8.5 ms minimum 0.1 to 0.2 ms Synchronized clocks. Then the wait must cover only the fiber and one peer stamp.
Card work for the order not measured separately 0.08 ms Nothing. A model of the card logic gives this value.
Execution on the host 135.51 ms 0.02 ms A batch must not wait for the certificate of the previous batch. The match itself is a few microseconds.
Batch close 4.62 ms 0.01 ms Close a batch without a fixed wait
Signatures and checks to FINAL 72.27 ms 0.35 ms FPGA cards at all sites sign in hardware. The card checks a signature in 0.154 ms (measured).
Full path approximately 226 ms approximately 0.6 ms

The targets are a design budget, not a measurement. The only measurement today is a silicon test of 30 Sep 2026. In it, the card went from an accepted order to FINAL in 0.33 ms, with four peer signature checks. That test did not contain the match or the network. We will publish the full path when we measure it.

Each part of the path on a log scale. A filled square is the last chain at p50, and a hollow square is the design target.
Figure 13. The last chain and the design target, part by part.

An artifact that we removed

On the policing chains, one metric compared the clock of the card with the clock of the host. It did not correct the offset between them. It showed approximately 21 s that was not real delay. We do not use that metric.

All latency on this page comes from inside the systems of the operator. It does not contain the network path of an outside client.

results matching ""

    No results matching ""