Re: kaio

In article <5opn7t$ojn@cssun.mathcs.emory.edu>, Clem Akins <clema@informix.com> writes >>I am sure there are people around who's done the comparison on how KAIO >performs >>against AIO VPs. Ten , fifteen per cent faster ? >>I am interested in HP , but any platform will do. > >YMMV! > >But I have personally done a benchmark on a large OLTP system on HP and saw a >10% reduction in average transaction time due to changing only KAIO. > Why? Since KAIO threads run in the same process as sqlexec threads (which in effect "generate" I/O requests). With KAIO there is - less context switching since less processes - less CPU cache flushes due to context switching and assocaited changing on MMU settings. - No need to wakeup AIO VPS which are currently awaiting I/O completion (which implies waiting fot the AIO VP to be rescheduled again). I'm confused unless KAIO is poorly implemented in some way.... What becomes the bottleneck when KAIO + many users is used? >I heard from an Informix tech who personally did his own BM and found the >performance >breakpoint was at about 3,000 concurrent (simulated with Tuxedo(?)) users, >after >which KAIO significantly hurt performance. Up until that point it helped, >but I >don't recall exactly how much. > >Aside from that, and the caveats mentioned in the release notes for certain >versions on HP, I don't know of any reason *not* to use KAIO of your platform >supports it. > >Good luck, >______________________________________________________ >Clem Akins (aka clema@informix.com) >Informix Software, Inc (Standard disclaimers apply) >International Technical Support >Last seen: Dodging raindrops in Wintery Auckland -- David Williams