Re: linux os cache and ocfs

Hi Mladen ,

Thanks for your answer. In fact your answer resulted another questions.

> It uses direct I/O only if you activate it and only if it is
> supported in kernel. You can activate direct I/O by setting the
> parameter filesystemio_options appropriately.

My setting for filesystemio_options is none . I was thinking to set it as async in order to enable aio till you had told me about the direct i/o. Whats the difference between direct i/o and async . Just ignoring this parameter , if i use ocfs , doesnt it open files by using o_direct? And doesnt this mean that the io to the file does not use any os buffer?

> Yes, you do. It's called "prefetch" and it does speed up access to the
> files that are mostly read only. Even more, Linux doesn't distinguish
> between data buffers in the buffer pool and page buffers. All modern
> Unix and Unix-like systems have blurred this distinction. Make no
> mistake, all Unix systems (and lookalikes) can support direct I/O, but
> doing so will usually cost you a penny or two. On Solaris, you should
> buy VxFS with QuickIO option. If you have 9.2.0.5 on Linux, you can use
> direct I/O against JFS file systems. Version 10 is scre...different.

I use emc disk array and 8gb . of cache on emc box , connected via fibre channel. I think when i use ocfs , this prefetching is not important.

> With swapping you can run many more simultaneous programs then without it.
> And, if you are not careful, swapping can easily kill the wabbit.

In fact i mean paging not swapping. As far as i know , unix systems does not use swapping but only paging . I really want to ask the reason of swap disk space usage. Consider a new process , when it first starts running , only the working set of this process must be in physical memory ( so thats why its called working set ) .When pages which are not in memory are needed , page fault occurs and data must be transferred from the disk. When this occurs , does the swap usage value in top display increase? Isnt it used in order to satisfy virtual memory paging ? Isnt it pagefile?(or vmstat si so values). In my case isnt my paging value high?

When the servers restarts , for a certain times , the swap usage is 0. But i have the processes running for instance oracle , does this mean that all the process executable is in memory?

Mladen , as you see i am little bit confused. Starting from the high level aio question , i at last have the low level virtual memory question.

I will be appreciated if you can clear these points.

Kind Regards,
hope

Mladen Gogala <gogala_at_sbcglobal.net> wrote in message news:<pan.2004.07.29.05.53.30.736464_at_sbcglobal.net>...
> On Wed, 28 Jul 2004 05:41:28 -0700, utkanbir wrote:
>
> > And most
> > importantly , i use ocfs which uses direct io ,
>
> It uses direct I/O only if you activate it and only if it is
> supported in kernel. You can activate direct I/O by setting the
> parameter filesystemio_options appropriately.
>
> > so do i get any
> > benefit of this linux disk cache? (since direct io does not use any os
> > buffer )
>
> Yes, you do. It's called "prefetch" and it does speed up access to the
> files that are mostly read only. Even more, Linux doesn't distinguish
> between data buffers in the buffer pool and page buffers. All modern
> Unix and Unix-like systems have blurred this distinction. Make no
> mistake, all Unix systems (and lookalikes) can support direct I/O, but
> doing so will usually cost you a penny or two. On Solaris, you should
> buy VxFS with QuickIO option. If you have 9.2.0.5 on Linux, you can use
> direct I/O against JFS file systems. Version 10 is scre...different.
>
> > Why does linux incrase the cache size too much and than
> > starts paging?
>
> Let me rephrase the answer given to a similar question by Miss Jessica
> Rabbit, the consort of Roger Rabbit: Linux is not bad, it was just
> written that way.
>
> > Whats the benefit of swapping ?
>
> With swapping you can run many more simultaneous programs then without it.
> And, if you are not careful, swapping can easily kill the wabbit.
Received on Thu Jul 29 2004 - 06:35:13 CDT