Re: request for enhancement: 2K -> 4K

In article <349E93E7.D0BB01E5@bru-co.be.dhl.com>, Walter De Smedt <wdesmedt@bru-co.be.dhl.com> writes >Jon Vemo wrote: >> >> At 11:49 AM 12/15/97 +0000, you wrote: >> }David Williams wrote: >> } >> }> In article <3494a300.1118117@news.xs4all.nl>, Reinoud van Leeuwen >> }> <reinoud@xs4all.nl> writes >> }> >On 12 Dec 1997 09:26:16 -0700, Pablo Sanchez <pablo@sgi.com> wrote: >> }> > >> }> >>Sybase Inc., >> }> >> >> }> >>I'd like to request that we increase the atomic page size >> }> >>from 2K to 4K. Today's systems I/O subsystems can handle >> }> >>the larger page sizes. It'd be nice to have pages greater >> }> >>than 2K for DSS type of applications. OLTP, I don't >> }> >>believe, would suffer. (Informix uses a 4K page size and >> }> >>they do pretty well...) with DSS app's, one has >> }> >>denormalization that tends to create fat rows... how about >> }> >>it? >> }> > >> }> >The disadvantage is that 4 K will be the minumum size for all >> }> >allocations, resulting in more wasted space in the database... >> }> >_______________________________________ >> }> >> }> With 2Gb (2048Mb) even wasting 20Mb is only 1%! Surely you don't >> }> size things that 1% more will mean you need to buy another disk? >> }> >> }> I come from tge Informix side and I'd like things similar to >> }> Informix's blobspaces. You have normal database spaces which have >> }> 2K or 4K pages depending upon the specific port. within blobspaces >> }> you then define how large a blobpage will be in units of either >> }> 2K or 4K. Big buffers (up to, I think 32K) are using to access >> }> blobspaces. >> }> >> }> Why can't we just define our own page sizes for each area of disk >> }> we use? Say in units of 2K or 4K depending upon the machine type? >> }> >> }> -- >> }> David Williams >> } >> }The paging functions in the Informix engine were complicated enough to >> }understand in v5.01 with only 2K pages, and I am unaware of how much it has >> }increased in the later versions, ie. 6+. I would think that this would not >> }be approved on the same grounds as you described above. Developing and >> }looking after this would simply be ugh! >> >> As I understand it, even in pre-6.x, the page size for Informix was platform >> specific. Most ports were done with a 2K page, however some which showed >> performance advantages with a larger page size, were ported as 4K. Sequent >> is one example. >> >> Jon > >We are currently facing severe I/O performance problems on several of >our HP9000 K2x0 SMP machines running HP-UX 10.10 and Informix Online I believe that HP-UX 10.20 fixes some problems in 10.01 and 10.10. >7.20. All these machines have a Nike HA disk array configured in RAID0/1 >or RAID1. This is clearly demonstrated by the long checkpoints during >operations (10-20sec). The problem is that the Unix SCSI driver has an >overhead of +/- 1ms per I/O operation. When a checkpoint needs to write >dirty buffers to disk containing 20MB of dirty data, and when you write >with a page size of 2k to raw devices, you have a total overhead of >10sec!! This has been verified with following command: > >time dd if=/dev/mem of=/dev/vginformix/rlvtest bs=2k count=10000 > >Real: 18,8sec > >An increased page size decreases the needed time dramatically. A >configurable page size would definately help. We don't mind weasting >some more disk space as a result of this. >[ A MIME text / x-vcard part was included here. ] > -- David Williams