[
Date Prev][
Date Next][
Thread Prev][
Thread Next][
Date Index][
Thread Index]
Re: HDF speed on netCDF API
- Subject: Re: HDF speed on netCDF API
- Date: Sun, 27 Mar 1994 14:27:22 -0700
> Organization: NOAA/PMEL
> Keywords: 199403182111.AA02131
Hi Steve,
> This is probably old news ... just FYI in case its not.
>
> We are porting FERRET to Solaris and decided it would be more convenient
> to use netCDF as embedded in HDF since we had to install HDF anyway.
>
> When we ran our benchmark it turned out that HDF has modified netCDF such
> that it is MANY TIMES SLOWER doing write operations. We didn't test
> extensively. (We just installed Unidata netCDF and relinked - much better
> now!)
Interesting, but it turns out that a performance comparison between the
HDF-encoded and XDR-encoded versions of the netCDF API is more complicated.
I think Chris Houck did a fairly reasonable job of tuning the HDF
implementation, and for some operations his code is faster. I ran the
"nctime" benchmark available in
ftp://ftp.unidata.ucar.edu/pub/netcdf/nctime.c on HDF3.3r2 and netCDF 2.3.2
on a SPARCstation 10 running Solaris 5.3. I've appended the results, which
write and read various hyperslabs of a 10x20x30x40 variable of each type:
byte, char, short, long, float, and double.
The very first write takes a long time (at least for our implementation)
because it writes fill values for all the variables. I've included the HDF
times, netCDF times, and the ratio of these times, so any numbers > 1.0 in
the ratio column indicate netCDF was faster by that factor. The results
show that our netCDF implementation that uses XDR is anywhere from 50 times
faster to 20 times slower, depending on the operation. These benchmark
times somewhat vary from run to run, so they aren't terribly accurate, but
they show places where both implementations could probably be tuned for
better performance. These results may be highly platform dependent and may
vary widely with other sizes of variables; I haven't tried "nctime" on
anything else. Also, I think there may be a more recent version of HDF out
by now.
--Russ
HDF netCDF HDF/netCDF
----- byte_var(10,20,30,40)
time for ncvarput 10x20x30x40 56.667 msec 1230.000 msec 0.05
time for ncvarget 1x1x1x1 0.661 msec 0.032 msec 20.66
time for ncvarget 10x1x1x1 5.882 msec 3.939 msec 1.49
time for ncvarget 1x20x1x1 13.333 msec 1.538 msec 8.67
time for ncvarget 1x1x30x1 16.667 msec 0.331 msec 50.35
time for ncvarget 1x1x1x40 0.775 msec 0.032 msec 24.22
time for ncvarget 10x20x1x1 83.333 msec 15.556 msec 5.36
time for ncvarget 10x1x30x1 123.333 msec 6.471 msec 19.06
time for ncvarget 10x1x1x40 5.882 msec 3.939 msec 1.49
time for ncvarget 1x20x30x1 250.000 msec 7.059 msec 35.42
time for ncvarget 1x20x1x40 12.222 msec 1.473 msec 8.30
time for ncvarget 1x1x30x40 0.853 msec 0.049 msec 17.41
time for ncvarget 10x20x30x1 2443.333 msec 56.667 msec 43.12
time for ncvarget 10x20x1x40 86.667 msec 15.556 msec 5.57
time for ncvarget 10x1x30x40 7.059 msec 4.545 msec 1.55
time for ncvarget 1x20x30x40 3.333 msec 1.846 msec 1.81
time for ncvarget 10x20x30x40 26.000 msec 20.000 msec 1.30
----- char_var(10,20,30,40)
time for ncvarput 10x20x30x40 53.333 msec 40.000 msec 1.33
time for ncvarget 1x1x1x1 0.389 msec 0.032 msec 12.16
time for ncvarget 10x1x1x1 4.848 msec 3.939 msec 1.23
time for ncvarget 1x20x1x1 11.176 msec 1.846 msec 6.05
time for ncvarget 1x1x30x1 15.556 msec 1.085 msec 14.34
time for ncvarget 1x1x1x40 0.661 msec 0.029 msec 22.79
time for ncvarget 10x20x1x1 66.667 msec 16.667 msec 4.00
time for ncvarget 10x1x30x1 90.000 msec 7.647 msec 11.77
time for ncvarget 10x1x1x40 5.152 msec 3.636 msec 1.42
time for ncvarget 1x20x30x1 226.667 msec 8.235 msec 27.52
time for ncvarget 1x20x1x40 11.111 msec 1.846 msec 6.02
time for ncvarget 1x1x30x40 0.775 msec 0.930 msec 0.83
time for ncvarget 10x20x30x1 1793.333 msec 50.000 msec 35.87
time for ncvarget 10x20x1x40 63.333 msec 17.778 msec 3.56
time for ncvarget 10x1x30x40 5.455 msec 5.152 msec 1.06
time for ncvarget 1x20x30x40 3.333 msec 2.154 msec 1.55
time for ncvarget 10x20x30x40 24.000 msec 20.000 msec 1.20
----- short_var(10,20,30,40)
time for ncvarput 10x20x30x40 120.000 msec 110.000 msec 1.09
time for ncvarget 1x1x1x1 0.545 msec 0.032 msec 17.03
time for ncvarget 10x1x1x1 5.455 msec 3.939 msec 1.38
time for ncvarget 1x20x1x1 8.824 msec 3.030 msec 2.91
time for ncvarget 1x1x30x1 11.111 msec 1.085 msec 10.24
time for ncvarget 1x1x1x40 0.545 msec 0.044 msec 12.39
time for ncvarget 10x20x1x1 60.000 msec 26.000 msec 2.31
time for ncvarget 10x1x30x1 93.333 msec 7.647 msec 12.21
time for ncvarget 10x1x1x40 5.152 msec 4.242 msec 1.21
time for ncvarget 1x20x30x1 176.667 msec 8.824 msec 20.02
time for ncvarget 1x20x1x40 8.824 msec 3.030 msec 2.91
time for ncvarget 1x1x30x40 0.775 msec 1.163 msec 0.67
time for ncvarget 10x20x30x1 1756.667 msec 63.333 msec 27.74
time for ncvarget 10x20x1x40 63.333 msec 28.000 msec 2.26
time for ncvarget 10x1x30x40 7.647 msec 8.235 msec 0.93
time for ncvarget 1x20x30x40 5.455 msec 10.000 msec 0.55
time for ncvarget 10x20x30x40 40.000 msec 73.333 msec 0.55
----- long_var(10,20,30,40)
time for ncvarput 10x20x30x40 203.333 msec 530.000 msec 0.38
time for ncvarget 1x1x1x1 0.775 msec 0.029 msec 26.72
time for ncvarget 10x1x1x1 5.882 msec 3.939 msec 1.49
time for ncvarget 1x20x1x1 11.176 msec 4.848 msec 2.31
time for ncvarget 1x1x30x1 12.222 msec 1.085 msec 11.26
time for ncvarget 1x1x1x40 0.545 msec 0.107 msec 5.09
time for ncvarget 10x20x1x1 93.333 msec 40.000 msec 2.33
time for ncvarget 10x1x30x1 126.667 msec 8.824 msec 14.35
time for ncvarget 10x1x1x40 6.471 msec 4.848 msec 1.33
time for ncvarget 1x20x30x1 186.667 msec 10.000 msec 18.67
time for ncvarget 1x20x1x40 10.000 msec 5.882 msec 1.70
time for ncvarget 1x1x30x40 0.930 msec 3.333 msec 0.28
time for ncvarget 10x20x30x1 2403.333 msec 80.000 msec 30.04
time for ncvarget 10x20x1x40 90.000 msec 50.000 msec 1.80
time for ncvarget 10x1x30x40 10.000 msec 24.000 msec 0.42
time for ncvarget 1x20x30x40 10.588 msec 40.000 msec 0.26
time for ncvarget 10x20x30x40 100.000 msec 356.667 msec 0.28
----- float_var(10,20,30,40)
time for ncvarput 10x20x30x40 226.667 msec 536.667 msec 0.42
time for ncvarget 1x1x1x1 0.661 msec 0.029 msec 22.79
time for ncvarget 10x1x1x1 7.059 msec 3.939 msec 1.79
time for ncvarget 1x20x1x1 11.111 msec 4.848 msec 2.29
time for ncvarget 1x1x30x1 17.778 msec 1.008 msec 17.64
time for ncvarget 1x1x1x40 0.700 msec 0.107 msec 6.54
time for ncvarget 10x20x1x1 86.667 msec 36.667 msec 2.36
time for ncvarget 10x1x30x1 123.333 msec 8.824 msec 13.98
time for ncvarget 10x1x1x40 5.882 msec 4.848 msec 1.21
time for ncvarget 1x20x30x1 253.333 msec 10.588 msec 23.93
time for ncvarget 1x20x1x40 11.111 msec 6.471 msec 1.72
time for ncvarget 1x1x30x40 1.163 msec 3.333 msec 0.35
time for ncvarget 10x20x30x1 2496.667 msec 76.667 msec 32.57
time for ncvarget 10x20x1x40 86.667 msec 50.000 msec 1.73
time for ncvarget 10x1x30x40 11.111 msec 24.000 msec 0.46
time for ncvarget 1x20x30x40 11.111 msec 36.667 msec 0.30
time for ncvarget 10x20x30x40 96.667 msec 356.667 msec 0.27
----- double_var(10,20,30,40)
time for ncvarput 10x20x30x40 436.667 msec 1060.000 msec 0.41
time for ncvarget 1x1x1x1 0.545 msec 0.032 msec 17.03
time for ncvarget 10x1x1x1 5.758 msec 3.939 msec 1.46
time for ncvarget 1x20x1x1 13.333 msec 7.647 msec 1.74
time for ncvarget 1x1x30x1 16.667 msec 1.085 msec 15.36
time for ncvarget 1x1x1x40 0.775 msec 0.176 msec 4.40
time for ncvarget 10x20x1x1 83.333 msec 66.667 msec 1.25
time for ncvarget 10x1x30x1 103.333 msec 11.111 msec 9.30
time for ncvarget 10x1x1x40 6.061 msec 5.882 msec 1.03
time for ncvarget 1x20x30x1 260.000 msec 14.444 msec 18.00
time for ncvarget 1x20x1x40 13.333 msec 10.000 msec 1.33
time for ncvarget 1x1x30x40 1.692 msec 5.455 msec 0.31
time for ncvarget 10x20x30x1 1943.333 msec 130.000 msec 14.95
time for ncvarget 10x20x1x40 83.333 msec 86.667 msec 0.96
time for ncvarget 10x1x30x40 14.444 msec 36.667 msec 0.39
time for ncvarget 1x20x30x40 22.000 msec 70.000 msec 0.31
time for ncvarget 10x20x30x40 166.667 msec 686.667 msec 0.24
