Re: ifx performace: linux vs NT

I agree with you. NT is most performance than Linux. Well, this is true until Informix has KAIO for Linux. Because: The cooked files aren't so cooked on NT. Try a onstat -g ioq on NT and Linux. How many KAIO VP's are? I guess you're going to find some KAIO ( they will be 2 ) on NT and no in Linux. I comprobed the difference between to use and no to use KAIO. It's 1:10 ( at least ) in the same unix box. You said NT is twice more then Linux. If you used KAIO on Linux will be 5 times more than NT, but you have to wait because KAIO on Linux isn't supported. Second. Unbuffered is not so unbuffered on NT. All transaction IO on NT pass through the cache. You can watch the Asyncronous Read Cache on Permance Monitor of NT Administrative tools. Francisco Huezo. I hope you can undertand me becouse my english is so bad. On Wed, 12 Jan 2000, Diego wrote: > > Palmer: thanks for your input in this isssue. Let me tell you about the > tests: > > I started with a set of queries against 5 tables: one with 150.000 rows, > other with 250.000 rows and the rest pretty small ones. > > First, forcing the engine to sequential scan the whole table. > Then, queries of a single table, with a filter condition on a indexed field, > with and w/o the order by clause. > And last, a series of joins with filters conditions with and w/o the order > by clause. > > In most of the tests the linux box was slower...twice I must say. For > example, the sequential scan took NT 1 min and linux 2. > > I was going to continue with a series of more elaborated tests, but with > this figures I panic and post my onconfig for help. I belive I must be doing > something wrong. Anyone with a bright idea? > > > > System: > > Linux RedHat 6.0 / Kernel 2.2.5-15 > Informix IDS 7.30 UC7 > PII 350 Mhz, single cpu > 128 RAM > > NT 4.0 > Informix IDS 7.30 TC 5 > same HW > > Cooked files on both, (unbuffered on NT). Statistics updated. > > > ----- Mensaje original ----- > De: richard palmer > Para: Diego > CC: Informix-list > Enviado: Martes, 11 de Enero de 2000 04:38 p.m. > Asunto: Re: special thanks and informix performance considerations > > > I took a look at the onconfig you attached. When you say you are > disappointed with the performance, what manner of testing did you > use ? Without an idea of your metrics, performance tuning is more a > shot in the dark as opposed to dealing with identified problem areas > and resolving these in an orderly manner. > I can state that (last I heard) you could not configure more than one > AIO VP on an NT system. This in itself can create a bottleneck > problem in terms of I/O. The onconfig I received showed NO entry > for AIO VPs, which defaults to one. > Configuration parameters that work efficiently on NT may not be > what works best on a unix based system like Linux. Try bumping > up the NUMAIOVPs to 1 per chunk or 2 per dbspace (opinions > vary on this, try both ?) to distribute the I/O. Also, try bumping up > CLEANERs to atleast one per chunk with a minimum of two. > have a nice day. > ---------------------------------------- Content-Type: application/octet-stream; name="onconfig" Content-Transfer-Encoding: quoted-printable Content-Description: ----------------------------------------