Well, as you are finding, this "simple" approach is causing you
problems.  The only thing I can add is to just be cautious about
relying on side-effects or presumptions about these multiple readers.
Because they could change at any time.  I've never considered such an
approach because I have always thought of a File Stream / Reader and
the actual file as a 1:1 relationship.  Anything else (except for 100%
read only) sounds dodgy.  You mention you sometimes going into append
mode.  But this would invalidate all the other read locks.

I agree that caches are not magic - they only work when on average,
access tends to cause clumps of common data.

If you were to pull apart the guts of database i wonder if you would
find multiple file readers.  My guess is no.

On Dec 26, 1:07 am, Eitan <[email protected]> wrote:
> Because it's a simple feature that I need and I don't want an overkill
> solution that I do not control its thread count, its memory
> consumption and cpu usage.
> It is just a simple component inside my highly performance server that
> needs to keep a single list of redo log in the disk if it takes to
> much memory which will be very light and won't impact the server
> performance. This occurs when the backup server it replicates to is
> temporarily down and the redo log starts to accumulate. There are many
> consideration which I can't really explain, but I do not wish to use
> any embedded or remote third party overkill product for this very
> simplified capabilities, like I've described above, unlimited single
> list of elements that can only be added at the end of the list,
> removed from the first and iterated over the list in read only mode.
>
> Any other generic product addresses a much more complex scenario and
> probably will perform worse because of that.
>
> I already implemented it and it works fine, but I am not sure about
> the multi reader over a file regarding the visibility issue and I was
> hoping that someone knows what kind of guarantee is there for the
> visibility of multiple RandomAccessFile inside the same process over
> the same file. (only one in rw mode and the rest in r mode)
>
> Thanks
>
> On Dec 25, 12:41 pm, Viktor Klang <[email protected]> wrote:
>
>
>
> > Why a file mate?
> > Why not Redis, Cassandra, MongoDB, SQLite whatnot?
>
> > On Fri, Dec 25, 2009 at 10:20 AM, Eitan <[email protected]> wrote:
> > > What about what I said above, using multiple FileChannel/
> > > RandomAccessFile to read from the same file (not writing, there's a
> > > single writer multi reader mutual exclusion protection over these
> > > FileChannel/RandomAccessFile access), all I am worried about is the
> > > visibility that updates the writer will do will be visible to the
> > > readers once done, FileChannel guarantee this and I suspect it must
> > > rely on this kind of guarantee from the RandomAccessFile that creates
> > > it though it is not documented, But according to the simple analysis I
> > > of the RandomAccessFile getChannel method I did, it seems that it does
> > > need to guarantee it in order for FileChannel to guarantee it.
>
> > > Since my implementation here is just an "unlimited" single list of
> > > objects that starts in the memory and once there's no room continue in
> > > the disk, I went for this solution and not using an over kill embedded
> > > data base. I only need to support appending objects at the end of the
> > > list, removing from the start and iterating over the list in read only
> > > mode. (Either from the start to the end or from the end to the start).
>
> > > Regarding the caching you suggested, I created caching mechanism and
> > > there are a few layers of abstraction over this file, but, I am
> > > talking here about the lowest level of the abstraction which access
> > > the file and that might occur concurrently and I want to remove the
> > > contention there as well. (for instance 10 threads currently iterating
> > > over the list all at a different location) besides the whole issue
> > > here is to move the data from the memory to the disk because I am out
> > > of memory space, so caching the 10g file is not an option, I only
> > > cache the end and the start of the list since they are accesses most
> > > frequently.
> > > The time spending ratio here is not the problem, there can be tens of
> > > threads concurrently iterating over the list and I want them to be
> > > able to iterate concurrently and not wait for each other. According to
> > > amdal's law even if the time spent on reading is just 10%, I can only
> > > accelerate this process by at most 10 so 100 threads iterating will
> > > work no faster than 10.
>
> > > Basically if I get the visibility guarantee then my multi
> > > RandomAccessFile in read mode over the file is good enough solution,
> > > let the disk and the OS do the caching of accessing the same file
> > > sections and keep my program more concurrent.
>
> > > Thanks
>
> > > On Dec 24, 11:45 pm, Christian Catchpole <[email protected]>
> > > wrote:
> > > > Is not the simplest solution to allow each thread to get a reference
> > > > to the *single* RandomAccessFile instance and simply synchronize on it
> > > > while reading (not forgetting to reset the offset into the file).
> > > > It's not going to be super performant, but how is it ever going to be
> > > > with multiple threads thrashing on one file?  If anything it might
> > > > reduce thrashing as you are controlling the points of file
> > > > contention.  What is the ratio of time spent reading the file to time
> > > > spent operating on the data?
>
> > > > If all these threads will in-fact thrash on the file it might only be
> > > > solved with:
>
> > > > 1. Better scheduling than "free for all" threads.
> > > > 2. a large OS level disk cache
> > > > 3. an in memory disk cache
> > > > 4. some kind of abstract layer over the file which does caching and
> > > > returns you "model objects" rather than being seen as a byte stream.
>
> > > > On Dec 25, 1:38 am, Eitan <[email protected]> wrote:
>
> > > > > Looking at FileChannel it has the following part in its documentation:
>
> > > > > The view of a file provided by an instance of this class is guaranteed
> > > > > to be consistent with other views of the same file provided by other
> > > > > instances in the same program. The view provided by an instance of
> > > > > this
> > > > > class may or may not, however, be consistent with the views seen by
> > > > > other
> > > > > concurrently-running programs due to caching performed by the
> > > > > underlying
> > > > > operating system and delays induced by network-filesystem protocols.
> > > > > This
> > > > > is true regardless of the language in which these other programs are
> > > > > written, and whether they are running on the same machine or on some
> > > > > other
> > > > > machine. The exact nature of any such inconsistencies are system-
> > > > > dependent
> > > > > and are therefore unspecified.
>
> > > > > This looks exactly like the guarantee I need, now I've looked into
> > > > > RandomAccessFile.getChannel() implementation and it holds a single
> > > > > member of FileChannel which is initialized on demand and the next
> > > > > calls will return the same instance.
>
> > > > > So the only way I see to create multiple FileChannel instances over
> > > > > the same file is to open multiple RandomAccessFile instances over that
> > > > > file and use getChannel() on each of these instances. Since
> > > > > FileChannel does provide this guarantee I tend to believe it relies
> > > > > upon a guarantee that RandomAccessFile provide, otherwise it would
> > > > > need some mechanism to identify that different FileChannel instances
> > > > > over different RandomAccessFile instances are actually over the same
> > > > > file in order to provide the specified guarantee.
>
> > > > > What do you think?
>
> > > > > On Dec 21, 9:33 am, Ben Schulz <[email protected]> wrote:
>
> > > > > > My guess is that Java IO takes advantage of any OS optimizations
> > > there
> > > > > > are, on another OS there may be a per-file-pointer cache or
> > > something.
> > > > > > That would obviously invalidate your assumptions. I suggest you look
> > > > > > into using asynchronous I/O which will get rid of any
> > > multiple-reader-
> > > > > > bottlenecks. This should also rule out third parties tampering with
> > > > > > the file (since you already hold the write lock).
>
> > > > > > With kind regards
> > > > > > Ben
>
> > > > > > On 21 Dez., 07:50, Eitan <[email protected]> wrote:
>
> > > > > > > I am trying to avoid locking the access to file for multiple
> > > readers
> > > > > > > can be multi concurrent readers and it will be a bottleneck.
> > > > > > > I've created a multi threaded test for it and it seems to work
> > > fine,
> > > > > > > but I am not still not confident about it. I was hoping to find
> > > some
> > > > > > > documentation about it which specifies the behavior.
>
> > > > > > > On Dec 20, 11:40 pm, Christian Catchpole <[email protected]>
> > > > > > > wrote:
>
> > > > > > > > Can you even get a second lock on a file that way?  Sounds
> > > dangerous.
> > > > > > > > I would just create a some kind of service that each Thread can
> > > access
> > > > > > > > that will synchronise disk access down to one RandomAccessFile.
>
> > > > > > > > On Dec 21, 2:48 am, Eitan <[email protected]> wrote:
>
> > > > > > > > > I want to use a file in a single process with single writer or
> > > multi
> > > > > > > > > readers mutual exclusion, for that purpose I use number of
> > > > > > > > > RandomAccessFile instances over the same file which are open 
> > > > > > > > > in
> > > read
> > > > > > > > > mode and are used concurrently. I also have a single
> > > RandomAccessFile
> > > > > > > > > open in "rw" mode which updates the file (not when it is being
> > > > > > > > > actively read).
> > > > > > > > > I do not close the Random Access File handles at any point.
>
> > > > > > > > > I wanted to know if the data which is being written to the 
> > > > > > > > > file
> > > with
> > > > > > > > > the RandomAccessFile in rw mode will always be visible to the
> > > readers
> > > > > > > > > after the the write is complete?
> > > > > > > > > (I am not using rws mode for performance issues, only for 
> > > > > > > > > these
> > > > > > > > > readers which are in the same process).
>
> > > > > > > > > Thanks
> > > > > > > > > Eitan
>
> > > --
>
> > > You received this message because you are subscribed to the Google Groups
> > > "The Java Posse" group.
> > > To post to this group, send email to [email protected].
> > > To unsubscribe from this group, send email to
> > > [email protected]<javaposse%2bunsubscr...@googlegroups
> > >  .com>
> > > .
> > > For more options, visit this group at
> > >http://groups.google.com/group/javaposse?hl=en.
>
> > --
> > Viktor Klang
> > | "A complex system that works is invariably
> > | found to have evolved from a simple system
> > | that worked." - John Gall
>
> > Blog: klangism.blogspot.com
> > Twttr: twitter.com/viktorklang
> > Code: github.com/viktorklang

--

You received this message because you are subscribed to the Google Groups "The 
Java Posse" group.
To post to this group, send email to [email protected].
To unsubscribe from this group, send email to 
[email protected].
For more options, visit this group at 
http://groups.google.com/group/javaposse?hl=en.


Reply via email to