My limited understanding, confirmed by profiler though, is that doing mmap IO cost you a copying bytes from mmaped virtual memory into heap VM. Just look into java.nio.DirectByteBuffer.get(byte[], int, int) . It happens several times to me - we saw hotspot in profiler on mmaped IO (yep, just in copying bytes!!), cache them in heap and we had hotspot moved after that. Good sample of heap cache for mmaped data is terminfos cache with configurable interval. Overal question is absolutely worth to think about.
On Thu, Sep 20, 2012 at 9:39 PM, Erick Erickson <erickerick...@gmail.com>wrote: > So I just had a curiosity question pop up and wanted to check it out. > Solr has the documentCache, designed to hold stored fields while > various parts of a requestHandler do their tricks, keeping the stored > content from having to be re-fetched from disk. When using > MMapDirectory, is this even something to worry about? > > It seems like documentCache wouldn't be all that useful, but then I > don't have a deep understanding here. I can imagine scenarios where it > would be more efficient i.e. it's targeted to the documents actually > being accessed rather than random places on disk in the fdt/fdx > files.... > > Thanks, > Erick > -- Sincerely yours Mikhail Khludnev Tech Lead Grid Dynamics <http://www.griddynamics.com> <mkhlud...@griddynamics.com>
