Disclaimer: the opinions expressed by the blog author do not necessarily represent the views of the Unidata Program Center.
Multithreaded TDS clients (1)
Jeff McWhirter was studying whether he could speed up IDV data access by multithreading on the client. He put together a test which fetched the same data over and over, with 1-10 client threads.
This is hitting motherlode TDS, which has 2 quad-core processors, running solaris x86 and ZFS. It asks for one field from an NCEP model file, stored in GRIB1 compression, so theres some amount of computation. Actual URL:
"http://motherlode.ucar.edu:8081/thredds/dodsC/model/NCEP/NAM/CONUS_80km/NAM_CONUS_80km_20090313_1200.grib1.dods?Temperature[0:1:0][0:1:18][0:1:64][0:1:92]";
19 * 65 *93 = 114855 floats = 459K bytes.
The result shows that you can't get more than 2.5x speedup in this particular case. I will try this on netcdf files, which should be only pure IO. Also interesting to see if the opendap stack causes significant overhead.
A more realistic test is to ask for successive time slices of non-cached data. These may be better opportunities for overlapping IO.
This is hitting motherlode TDS, which has 2 quad-core processors, running solaris x86 and ZFS. It asks for one field from an NCEP model file, stored in GRIB1 compression, so theres some amount of computation. Actual URL:
"http://motherlode.ucar.edu:8081/thredds/dodsC/model/NCEP/NAM/CONUS_80km/NAM_CONUS_80km_20090313_1200.grib1.dods?Temperature[0:1:0][0:1:18][0:1:64][0:1:92]";
19 * 65 *93 = 114855 floats = 459K bytes.
The result shows that you can't get more than 2.5x speedup in this particular case. I will try this on netcdf files, which should be only pure IO. Also interesting to see if the opendap stack causes significant overhead.
A more realistic test is to ask for successive time slices of non-cached data. These may be better opportunities for overlapping IO.