libbls-dev-list team mailing list archive

[Alexandros Frantzis <alf82@xxxxxxxxxxx>]

> >
> > So it seems that the avoidable overhead is that of the use of memcpy.
> > An idea around it is to have a new API function eg
> > bless_buffer_read_foreach() that will act just like segcol_foreach().
> >

I implemented a first version of such a function (not yet uploaded) and
ran the benchmarks again. The results were very encouraging for both the
file and memory cases: we are as efficient as plain-read for the memory
case and (very) slightly less efficient than mmap for the file! I have
attached the results of the benchmarks.

Note that these results were taken using a bless_buffer using a 1000
segments (only a single segment was used in the previous tests). The
good performance that is exhibited in this case makes the results even
more encouraging. More tests are needed to check how read performance
scales with the number of segments. My guess is that performance will
drop significantly when we have many segments smaller than the page size
(for the file case, at least).

> >
> > The question is why do we have a 40% overhead in bless_buffer_read(),
> > considering that we just do a memcpy, too?
> >

That still remains to be investigated...

-- 
Alexandros

25 0.090000 0.056667 0.050000 0.060000 0.073333 0.070000 0.036667 0.036667 0.053333 
50 0.173333 0.116667 0.103333 0.110000 0.160000 0.133333 0.073333 0.080000 0.100000 
75 0.260000 0.166667 0.156667 0.166667 0.240000 0.200000 0.110000 0.120000 0.150000 
100 0.340000 0.223333 0.210000 0.223333 0.310000 0.273333 0.150000 0.153333 0.200000 
125 0.433333 0.276667 0.260000 0.273333 0.396667 0.340000 0.193333 0.190000 0.250000 
150 0.516667 0.333333 0.306667 0.326667 0.480000 0.396667 0.230000 0.230000 0.300000 
175 0.606667 0.386667 0.366667 0.383333 0.553333 0.466667 0.263333 0.270000 0.350000 
200 0.690000 0.450000 0.413333 0.433333 0.636667 0.540000 0.306667 0.310000 0.403333 
225 0.773333 0.503333 0.470000 0.483333 0.720000 0.603333 0.343333 0.350000 0.450000 
250 0.883333 0.553333 0.520000 0.546667 0.793333 0.680000 0.383333 0.386667 0.496667 
275 0.950000 0.616667 0.570000 0.600000 0.866667 0.733333 0.423333 0.423333 0.550000 
300 1.066667 0.673333 0.616667 0.656667 0.956667 0.830000 0.460000 0.460000 0.603333 
325 1.140000 0.723333 0.676667 0.703333 1.030000 0.920000 0.496667 0.500000 0.650000 
350 1.233333 0.783333 0.726667 0.763333 1.103333 1.013333 0.540000 0.540000 0.700000 

plot "bench_read.data" using 1:2 with linespoints title "file:bless_buffer_read",\
     "bench_read.data" using 1:3 with linespoints title "file:bless_buffer_read_foreach",\
     "bench_read.data" using 1:4 with linespoints title "file:read",\
     "bench_read.data" using 1:5 with linespoints title "file:mmap",\
     "bench_read.data" using 1:6 with linespoints title "file:mmap+memcpy"


pause -1

plot "bench_read.data" using 1:7 with linespoints title "mem:bless_buffer_read",\
     "bench_read.data" using 1:8 with linespoints title "mem:bless_buffer_read_foreach",\
     "bench_read.data" using 1:9 with linespoints title "mem:plain",\
     "bench_read.data" using 1:10 with linespoints title "mem:memcpy"


pause -1