David Daney wrote:
Ian Rogers wrote:
Hi,
please give your comments on the attached patch. It tries to reduce
the size of char[] for strings used to hold numbers. It changes
Float/Double equals to use bit based comparisons rather than
division. It increases the use of valueOf methods. It adds a cache of
values from -128 to 127 for Long. It adds a cache of the values of
zero and one to Float and Double.
The string size is an estimate. For decimal numbers it will divide
the value repeatedly by 8, causing the string length to be over
estimated by a character for values like 999. This string size is
still better than the current estimate of 33 characters. It also
avoids the use of division (shifts are used) and/or lookup tables.
I would like to know your motivation for doing this. Do you have any
evidence that this will reduce memory usage and speed up real
applications?
That said, in our gcj-3.4 based application, we had to create a cache
of Integers because we were creating large numbers of them all with a
small set of values.
So in principle this could be a good approach, but I don't know if we
can assume that there is universal benefit from a patch like this.
Can you point to any benchmarks where this helps?
Thanks,
David Daney
Hi David,
I'm having a crack down on wasted memory in the Jikes RVM.
For DaCapo fop (single iteration) there are 270 and 977 occurrences of
Double 0 and 1 and 20 occurrences of other Doubles. On the other hand
DaCapo bloat has very few 0 and 1 values. My motivation to cache these,
other than fop, is that they exist as bytecodes (fconst0/fconst1 and
dconst0/dconst1, although I'm ignoring fconst2 and dconst2). We already
cache Integers in the intCache. I do extend this concept to Long, as is
done in OpenJDK, and to Float and Double.
Currently we always allocated 33 char arrays to hold the string value,
this is 4.625 the size of the minimum object in the RVM. In the case of
a single character string, 18.86% of Integer strings in DaCapo bloat,
this code doesn't allocate any char arrays. For other integers the char
array is reduced to either the exact or (20% of the time for decimal
values) 1 character longer char arrays. This is at the cost of up to 32
compares, branches and shifts. For DaCapo bloat a little under 50% of
integer strings created are for values between -128 and 127.
So the trade offs in the code are, slower Float/Double valueOf code, but
fewer Float and Double objects (hopefully improving GC). A small time to
calculate string sizes vs smaller strings and less GC pressure.
For the Jikes RVM we measure performance 4 times a day [1], I introduced
this patch in r14113 and there are no peaks or troughs that appear at
this point. Given the patch is performance neutral but saves memory
(although not improving GC performance for the RVM markedly) I think
it's worth including. GC is less than 6% of execution time, so time
saved may be difficult to measure in the bigger picture (unless it
pushes you under or over a particular threshold).
Regards,
Ian
[1]
http://jikesrvm.anu.edu.au/cattrack/results/rvmx86lnx32.anu.edu.au/perf/3437/performance_report