> I think that's the bit you're missing. 8 bits I'm missing, actually. BA-ZING!
64-bit implementations of Java align on 8-byte boundaries. A 2-byte value plus a 4-byte value will be packed into 8 bytes. An 8-byte double and a 4-byte int will add 4 useless bytes with alignment, giving 16 bytes.. Including 16 bytes per object, Compact v.s. Full is 24 bytes v.s. 32 bytes. Since there are 2 unused bytes in the Compact object, the 16-bit 'char' counter might as well be a 31-positive-bit counter. Similarly, the integer counter in FullRunningAverage could be a long instead of an int. For comparison, an empty String is 40 bytes, or 5 8-byte words. On Fri, Mar 25, 2011 at 2:33 AM, Sean Owen <sro...@gmail.com> wrote: > A char is a 2-byte unsigned value in Java, not 1-byte signed -- I > think that's the bit you're missing. > > Ignoring object overhead, there are 2 (char) + 4 (float) = 6 bytes of > payload in CompactRunningAverage. There are 4 (int) + 8 (double) = 12 > bytes of payload in FullRunningAverage. > > I don't know what you mean that 'count' should be an int. > > I do agree that it would be more memory-efficient to make a sort of > array of these things, in cases where you have a dense sequence of > them. However the use case you cite isn't one of those -- those diffs > are sparse. You can add into the object the row/col position but then > that just eats up the space savings. > > On Fri, Mar 25, 2011 at 9:19 AM, Lance Norskog <goks...@gmail.com> wrote: >> The CompactRunningAverage class is >> 1) not compact, and >> 2) not coded right. >> >> The code using 'count' seems to think it is a short. The code assumes >> that shorts top out at 65535, >> but shorts are signed and thus top out at 32767. >> 'count' is declared as a 'char' so it will never reach 128, let alone 65536. >> The 'count' field uses 4 bytes so it should be an int. >> Any object has 16 bytes. The CompactRunningAverage has (at least) 24 >> bytes. FullRunningAverage has (at least) 32 bytes. >> >> Something that supplies compact running averages should directly >> allocate and manage an array of shorts and an array of floats. >> MemoryDiffStorage is the only use of CompactRunningAverage in the code >> base. It can use "hundreds of thousands" of these counters. >> It also needs 8 bytes per counter for each object reference, v.s. 4 >> bytes per array index. >> >> -- >> Lance Norskog >> goks...@gmail.com >> > -- Lance Norskog goks...@gmail.com
