Hi Martin, Thank you very much for your clear and detailed answer. You are right, you said responsive in that podcast, sorry. But I would have done the maths anyway :)
Michaël On Saturday, September 28, 2019 at 3:22:12 PM UTC+2, Michaël REMOND wrote: > > Hello, > > I know about Martin Thompson excellent work since a good time now, and > recently I wanted to better understand some queuing theory he discussed in > the Arrested Devops podcast. > > He gave the example of a service having an average response time of 100ms, > and this service receives on average 9 requests/second. > If the service response time is divided by 2 (50 ms), Martin said that the > service becomes 20x times more "reactive". > > I was not sure what he meant exactly by reactive, but I tried to > understand where these 20x came from. > > Here is what I came to, and I wanted that some experienced people confirm > to me if my reasoning is correct or not. > > I assume that Martin is considering the M/M/1 queue model. Then the theory > says that on average, the *waiting time* (not the sojourn time) is: *ρ*/( > *μ* − *λ*). > With the real numbers, we have: > - case 1 (service time = 100ms): waiting time = (9/10) / (10 - 9) = 0.9s > = 900ms > - case 2 (service time = 50ms): waiting time = (9/20) / (20 - 9) = 0.04s > = 40ms > > And then 900 / 40 ≈ 22. So I think that when Martin is saying that the > system is more reactive, he is talking about the waiting time. > > So the first take-away for me, is that our systems should not run at a > utilization bigger than 80% (see > https://www.johndcook.com/blog/2009/01/30/server-utilization-joel-on-queuing/#comment-13511 > ). > > So my next question is: should I consider this result true for our CPUs > and memory as well? > > I am particularly interested in this question, because I read in the > Google SRE Book that for optimal resource utilization (and thereafter for > optimal costs), they try to make their CPUs almost always full of tasks. > > Thank you in advance for your responses and thoughts on this subject. > > Michaël > -- You received this message because you are subscribed to the Google Groups "mechanical-sympathy" group. To unsubscribe from this group and stop receiving emails from it, send an email to mechanical-sympathy+unsubscr...@googlegroups.com. To view this discussion on the web, visit https://groups.google.com/d/msgid/mechanical-sympathy/bbc8b35e-8892-4360-a3d6-165344f50782%40googlegroups.com.
