Return to BSD News archive
Newsgroups: comp.os.386bsd.questions Path: sserve!newshost.anu.edu.au!munnari.oz.au!news.Hawaii.Edu!ames!saimiri.primate.wisc.edu!news.doit.wisc.edu!psl.wisc.edu!128.104.200.15!ram From: ram@xor.epi.wisc.edu (Ram Bhamidipaty) Subject: Re: VLB IDE cache controller In-Reply-To: rand@cs.UND.NoDak.Edu's message of 22 Sep 1993 16:18:13 -0500 Message-ID: <RAM.93Sep23132553@xor.epi.wisc.edu> Sender: news@pslu1.psl.wisc.edu (USENET News System) Organization: Physical Sciences Lab, UW-Madison References: <27mf1g$ajt@news.u.washington.edu> <RAM.93Sep21145412@xor.epi.wisc.edu> <RAND.93Sep22161811@agassiz.cas.und.NoDak.Edu> Date: 23 Sep 1993 18:25:53 GMT Lines: 45 >>>>> "Douglas" == Douglas K. Rand <rand@cs.UND.NoDak.Edu> writes: >>> I just upgraded my old ISA IDE controller to a VLB cache version. >>> However, after some thorough test, I notice that the cache >>> controller doesn't seem to improve the I/O performance at all. Ram> I would expect almost no benefit from a caching controller, since Ram> the cache system in 386bsd is already doing that job. [...] but Ram> an IDE controller can only handle two disks, which can be handled Ram> by the ISA bus. Douglas> I've got a ... VLB caching controller with 4 Mbytes Douglas> controller (20 Mbytes main memory) and I see a signfigant Douglas> difference with the cache. Here are some simple performance numbers Douglas> for just the linking phase of building a new kernel: Douglas> First time Second time Douglas> cache disabled: 16.41 secs 16.21 secs Douglas> cache enabled: 9.20 secs 3.28 secs Point taken. But I can't help but wonder what would happen if (a) the 4meg was taken from the controller and put in main memory [if it was possible to "move" the memory] and (b) the kernel buffers were increased to 4meg. Not that I know how to do this. The kernel has all the information that the controller does, so in theory it should be possible to have the kernel duplicate and performance increase that one can gain from a caching controller. I have a vague recollection of someone saying that the kernel will use all available main memory for disk buffering if no other process is using it...is this true? Or is there some limit on how much buffer space the kernel uses for disk io. If the first statement is true then there might be some in-efficiency in the kernel buffer caching algorithms. If one needs to configure the kernel disk io buffers when building the kernel then it would be an interesting experiment to see how close the two performance figures come when the kernel is given an equal amount of cache memory. -Ram -- -------------- Ram Bhamidipaty use this ->ram@epidat.epi.wisc.edu Department of Preventive Medicine or this->ram%epidat.decnet@macc.wisc.edu University of Wisconsin, Madison Ignore the header. It might be wrong.