Article Discussion
Gathering Scattered I/O
Summary: Have your cake and eat it, too, with STL extensions. In this chapter extract from his latest book, Matthew Wilson shows you how to take full advantage of the STL Iterator abstraction, without sacrificing block-transfer efficiency of Scatter/Gather I/O memory.
2 posts.
The ability to add new comments in this discussion is temporarily disabled.
Most recent reply: September 28, 2007 11:48 AM by giovanni
    Frank
     
    Posts: 135 / Nickname: fsommers / Registered: January 19, 2002 7:24 AM
    Gathering Scattered I/O
    September 18, 2007 10:30 AM      
    Have your cake and eat it, too, with STL extensions. In this chapter extract from his latest book, Matthew Wilson shows you how to take full advantage of the STL Iterator abstraction, without sacrificing block-transfer efficiency of Scatter/Gather I/O memory.

    http://www.artima.com/cppsource/scattered_io.html

    What do you think of Wilson's take on Scatter/Gather I/O?
    • Daniel
       
      Posts: 1 / Nickname: djberg96 / Registered: September 21, 2004 2:45 PM
      Re: Gathering Scattered I/O in C++
      September 27, 2007 2:47 PM      
      Isn't this somewhat built into the Windows API now (since Windows 2000) with ReadFileScatter() and WriteFileGather()?
    • giovanni
       
      Posts: 1 / Nickname: gpderetta / Registered: September 28, 2007 4:22 AM
      Re: Gathering Scattered I/O in C++
      September 28, 2007 11:48 AM      
      std::deque is usually implemented exactly like a scatter/gather buffer range: a vector of pointers to (fixed size) buffers, but still it manages to get
      good performance.

      A known trick is to use segmented iterators (See the paper "Segmented Iterators and Hierarchical Algorithms" by Matt Austern): a special iterator protocol is used to segment a range in a ranges of ranges, defined by simpler iterators. Such protocol can be used by standard algorithms to iterate more efficiently over hierarchical data structures like, for example, deques. This way you have maximum efficiency without sacrificing transparency: algorithms and containers are still completely orthogonal. Also note that if implemented correctly you can have more than one level of segmentation.

      Unfortunately this protocol is not made available to the users by standard libraries implement it, so users cannot take advantage of it for their own containers. Coupled with the fact that you can't use a deque for vectored IO (there is no way to know where the contiguous buffers begins and end, i.e. you do not have access to the protocol itself), it means that the user must roll its own solution.

      I've implemented my own deque with segmented iterators and wrappers for some common standard algorithms that take advantage of this property.

      I've used successfully such a deque with s/g io: my own I/O abstraction was capable of being fed with any container type and used traits classes to detect whether the container was capable of direct I/O (for example a plain char array, or a std::vector<char>) or had segmented iterators over direct I/O subranges (my own deque).

      The framework was very flexible and most importantly, I/O was completely decoupled from the buffer abstraction. It could be easly extended to work with any third party container.

      BTW, my deque is a random access range, thus all blocks must be of same size. But if you weaken the requirements to bidirectional range, you can have an hibrid list/vector like structure that behaves like a bidirectional range, whose block view support splice, and its segmented iterators support random access within the block. Someday I'll get around implementing it. It should be easier than a deque.

      You can find my library at libstream.sourceforge.net. I haven't been working on it for ages (my current work is not network programming unfortunately), but someday I will clean it up.

      gpd