tag:blogger.com,1999:blog-6977755959349847093.post7805141447886055134..comments2024-03-01T21:53:15.921-08:00Comments on The Pith of Performance: Prime Parallels for Load BalancingNeil Guntherhttp://www.blogger.com/profile/11441377418482735926noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-6977755959349847093.post-46733702013802032772010-07-06T13:00:54.604-07:002010-07-06T13:00:54.604-07:00Glad you like it. :)
re: diagnosing a real load-b...Glad you like it. :)<br /><br />re: diagnosing a real load-balancer<br /><br />It would be interesting to try it. The only drawback I can see is the weak logarithmic convergence, which implies that many data samples would be required. But that's not a show stopper.<br /><br />We see from this analysis that Amdahl scaling A(p) is worse than G(p) because it corresponds to severe skewing between only two possible processor subsets: all (parallel with m=p) or one (serialized with m=1).Neil Guntherhttps://www.blogger.com/profile/11441377418482735926noreply@blogger.comtag:blogger.com,1999:blog-6977755959349847093.post-41827987647205511342010-07-06T11:01:38.067-07:002010-07-06T11:01:38.067-07:00This is fascinating. In other words, when the loa...This is fascinating. In other words, when the load balancing works properly (uniform distribution of "p"), the logarithmic equation will describe the speedup. <br /><br />Since you are using the Taylor sequence, which works in both directions, this can also be used for the inverse problem, i.e., to diagnose the load-balancer operation: if the operation does not scale as you describe in (10), it means that the "p" distribution is not uniform, and the load-balancing needs to be scrutinized.<br /><br />I was also wondering how low in "p" the (10) will hold: the asymptotic limit at p == 0 is not the problem, as we can't have fewer than 1 server running, but the Taylor decomposition may cause an accuracy issue at low numbers.AlexGilgurhttps://www.blogger.com/profile/01961676712058514905noreply@blogger.com