Cinnober just announced the publication of an in-depth whitepaper on latency on their web site:
The paper contains a much more detailed explanation for the items discussed in my previous post.
Taking into account that the figure (286 mks for door-to-door and 138 mks for business logic) is supported by the description of what was measured and how, it looks quite credible.
What is important is that even at this level of detail there is some room for interpretation of the door-to-door latency. Quote (page 6): “means measuring the time which elapses between a request being received at the AP and a corresponding response arriving
back at the same AP after being processed by the matching engine.”
What should one consider to be the time when the AP received a request? Is it the time when the data appeared in the socket or the time when the AP actually started processing the request? Were these orders and quotes submitted against an empty book? Were any
of these orders and quotes filled, acknowledged or rejected? At the microsecond scale, it makes quite a difference.
The paper contains a very good explanation of the queue theory and the service time calculation. One caveat, however, is: the obtained service time (98 mks – see page 13) is actually 4 times smaller than the real processing time, as there are 4 trading gateways
that simultaneously process incoming messages. The actual processing time was 392 mks.
Overall, I would highly recommend this excellent paper to anyone interested in execution venue latencies!