Systems and low latency
Two systems: one has a mean of 2 μs and a p99.9 of 2 ms, the other a mean of 5…
Two systems: one has a mean of and a p99.9 of 2 ms, the other a mean of and a p99.9 of . Which do you want on a market-data path?
- AThe first, since it is faster on average
- BThe second, because its tail is far tighter
- CEither – the difference is only one in a thousand messages
- DThe first, then add a retry when it is slow
Try it
Pick an answer
Next free question: Compute 48 × 52 in your head.
Show the answer and worked solution
Answer: B – The second, because its tail is far tighter
Take the second. A market-data path is judged on jitter rather than on the average, because slow paths are correlated with bursts and bursts are exactly when the messages matter: a two-millisecond stall lands during the news event you were positioned for. A predictable five microseconds is worth far more than an unpredictable two. "Only one message in a thousand" is precisely the wrong way to think about it, since that message is disproportionately the one that counted, and a retry does not help because by the time you know the path was slow the opportunity is gone and you have added load to a system already struggling.
Worked solution
- Formula
- Substitute
- Solve
- Which is when the messages matter most.
- Answer
Sanity check. A retry does not help, because by the time you know the path was slow the opportunity is gone.
- A. Its tail is a hundred times worse, and the tail lands during the bursts that follow news.
- B. Correct. Jitter is the target: a predictable 5 µs beats an unpredictable 2 µs.
- C. That one message is disproportionately the one that mattered.
- D. By the time you know it was slow, the opportunity has gone and the retry adds load.
Takeaway: A trading system is judged on its tail, not its mean.
Report an issue with this question
Learn the method
Latency, the memory hierarchy and why the tail is the number
Of the kind reported at
Optiver, Citadel Securities, Hudson River Trading, Jump Trading
More systems and low latency questions
- An L1 cache reference costs about 1 ns and a main-memory reference about 100 ns.Foundation · Free solution
- A burst on the feed, part 3 of 4Advanced · Premium
- Your load generator sends the next request only after the previous response arrives.Expert · Premium
- You sort 1,000 latency samples in ascending order.Foundation · Premium
- A timestamp counter ticks at 3 GHz. How many ticks elapse in one microsecond?Foundation · Premium
- A burst on the feed, part 1 of 4Foundation · Premium