[Larceny-users] GC options
Felix Klock
felixluser at pnkfx.org
Wed Jan 24 14:34:23 EST 2007
On Jan 24, 2007, at 12:45 AM, Felix Klock wrote:
> The -reorganize-and-dump option will generate a new heap file with
> the objects partitioned into two areas: one for the codevectors, and
> another for everything else. This can yield significant performance
> gains on IA32, because mutations to data should not affect the code
> cache after such partitioning is performed. (I will probably have
> concrete numbers on the benefits of doing this in the next day or
> two.)
Here are some concrete numbers illustrating the potential effects of
using -regorganize-and-dump.
The NONDUMP column is the standard set of larceny benchmarks, running
on a development version of the runtime. The DUMP column is the same
set of benchmarks, but now the benchmark compilation includes a step
where we load the benchmark, dump the interactive heap, and then
reorganize-and-dump the generated heap image. Running the benchmark
is then just a matter of loading the reorganized heap
In several cases, reorganize-and-dump yields a speed up of 5 to 20%,
but it not an absolute win. I'm a bit surprised by how much
performance degrades in some cases (e.g. boyer, diviter, divrec,
scheme); I do not have an immediate reason for it.
Anyway, the point is that Sven may want to try using -reorganize-and-
dump, just to see if it helps.
-Felix
"time (ms)" "time (ms)"
DUMP NONDUMP
boyer 1038 676
browse 3399 3208
cpstak 1576 1911
ctak 3657 3728
dderiv 1819 2096
deriv 1332 1568
destruc 1075 1138
diviter 2023 1384
divrec 2651 1968
puzzle 1535 1475
takl 1072 1061
triangl 1954 2873
fft 1493 1698
fib 2430 2425
fibfp 3021 3324
mbrot 2679 2967
pnpoly 1761 1805
sum 385 384
sumfp 2080 2468
tak 1708 1697
ack 96 95
array1 388 416
cat 356 355
string 640 644
sum1 893 779
sumloop 508 495
tail 554 565
wc 216 222
conform 2658 2664
dynamic 1051 1039
earley 2341 2449
fibc 3758 3773
graphs 1691 1815
lattice 2208
matrix 2289 2364
maze 7589 7648
mazefun 1598 1444
nqueens 1560 1621
paraffins 4190 4370
peval 1677 1720
primes 10400 10477
ray 5024 5168
scheme 2525 1868
simplex 5911 6252
slatex 396 394
perm9 986 1075
nboyer 1871 1933
sboyer 1807 1788
gcbench 1980 2075
parsing 8738 9120
gcold 21354 25098
More information about the Larceny-users
mailing list