lkml.org 
[lkml]   [1998]   [Jun]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: Thread implementations...
    > > True, aio_* with page aligned buffers and vm hacks amount to the same
    > > thing anyhow, but I don't know of anyone actively working on the mm
    > > system in this area.
    >
    > AFAIK, there is indeed nobody working on this. If people
    > actually want this feature, I'll add it to the TODO-list
    > on http://linux-mm.home.ml.org/...

    People are going to give you lots of reasons why this isn't a good idea. I'm
    going to say it has implication outside of networking that make it look very
    attractive.

    There was some dialog between Doug Ledford and myself on linux-smpdev about
    this, and whilst I originally argue it wasn't a good idea now have to
    concede it probably is.

    A couple of reasons, access patterns like:

    read(from,buf,len);
    write(from,buf,len);

    can me optimized to DMA to memory, twiddle some pages for the process, and
    avoid the costly copy to/from buffer cache that presently limits disk IO
    throughput (preserving the buffer cache using COW).

    At the moment, if I want to copy a 50MB file from one device to another,
    assuming I have lots of memory, the data first gets loaded into the buffer
    cache (say, via DMA), then that is copied to userspace, which then does a
    write and userspace gets copied to buffer-cache, which is then written to
    the destination device.

    Now, thats horrendously expensive. If we allowed mm tweaks, it gets DMA's
    into memory, and DMA'd out again, the buffer cache gets to share pages with
    userspace and when dirtied we can use COW or discard the buffer-cache pages
    when memory is low.

    I really think something like this would rock. Especially as FCA raid arrays
    and stuff are appearing, with pretty scary IO throughput....

    If you follow this argument through, and make minor modifications to
    userland code (such as 'cp'), you can also make this pretty click when
    copying to/from NFS.



    -Chris

    -
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to majordomo@vger.rutgers.edu

    \
     
     \ /
      Last update: 2005-03-22 13:43    [W:0.020 / U:4.044 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site