tag:blogger.com,1999:blog-3822757291061444396.post3558929320653505236..comments2020-02-23T04:43:07.427-06:00Comments on From the Canyon Edge: A Statistical Analysis of Potential PowerNap Energy SavingsDustin Kirklandhttp://www.blogger.com/profile/12464590128908584782noreply@blogger.comBlogger3125tag:blogger.com,1999:blog-3822757291061444396.post-62385863138871332422009-08-20T15:45:37.746-05:002009-08-20T15:45:37.746-05:00Parameters:
1) Number of hosts
2) Number of cl...Parameters:<br /> 1) Number of hosts<br /> 2) Number of clients per host<br /> 3) Time period of observation: ex. one week,one month.<br /> 4) Probability distribution for wait time since last<br />client created -- first moment is average period between client births, so equivalent to an average birth rate. <br /> 5) Probability distribution for lifetime of active host. First moment is is average lifetime, sort of equivalent to death rate..but its complicated because its implicitly depends on parameter 3).<br /><br />So pick sane distributions for 4 and 5. Use some intuition concerning expected work load...just make sure the limits are sane. If you have any real world logs you can use those logs to model these distributions. Don't worry too much about it for the first sims. Once these are in place, you use them to randomly pick client lifetimes and connection waits to let the number of clients dynamically grow instead of enforcing a number of clients.<br /><br />Start with 10 hosts, with 4 clients per host...run multiple sims for the specified time window and build up stats on the behavior.<br /><br />Increment the number of hosts and repeat. Keep doing that. Hopefully you'll see the power savings reach an obvious limiting trend after a few hosts.<br /><br />Go back increment the time window and do it again. Hopefully you'll see the power savings reach an obvious limiting trend as a function of time window for long times.<br /><br />Go back and choose 3 and 5 clients per host. See if the limits have changed much. Hopefully they haven't.<br /><br />Go back and do and make a small tweak to the distributions and see if it perturbs the long time, many hosts limiting trends. Hopefully small changes in workload are not magnified in the power savings limits...that would indicate highly nonlinear behavior..and generalized observations would seldom be a good measure for a particular workload. I'm not expecting to see this as there's nothing obvious in setup that would drive a feedback loop.<br /><br />The key is running "enough" sims at the same host number settings using those probability distributions. You get some mean and variance stats on your power savings. <br /> <br />-jefJef Spaletahttps://www.blogger.com/profile/11439754449677675460noreply@blogger.comtag:blogger.com,1999:blog-3822757291061444396.post-3344604870138708832009-08-20T11:01:04.101-05:002009-08-20T11:01:04.101-05:00Jef-
I've been thinking about this. What do ...Jef-<br /><br />I've been thinking about this. What do you propose are the parameters of such a model?<br /><br />Hosts, Guests-per-Host, Average-Guest-Life, "Churn", etc.?<br /><br />I'm having trouble quantifying the configurables.<br /><br />:-DustinDustin Kirklandhttps://www.blogger.com/profile/12464590128908584782noreply@blogger.comtag:blogger.com,1999:blog-3822757291061444396.post-28377076068346961612009-08-20T10:58:01.399-05:002009-08-20T10:58:01.399-05:00Quick once over and it looks pretty reasonable.
I...Quick once over and it looks pretty reasonable.<br /><br />If you wanted to geek out, you could probably model the expected savings in a dynamic system using an exponential probability distribution function for the times between virtual machine creation and again for virtual machine termination. You could even probably run a few simulations of a dynamic cloud using those probability functions in pretty short order. This would give you an idea of how the greedy algorithm interacts with the powernap over some time window.Jef Spaletahttps://www.blogger.com/profile/11439754449677675460noreply@blogger.com