26 May 2015
16 March 2009
A more realistic test is to ask for successive slices of some data, and see if multithreading speeds things up. So I made a test that makes data requests like send:
Each request is 1 Mbyte. These are grib2 files, which may be jpeg2000 encoded (im not sure in this particular file).
So each thread makes 20 requests of successive slices, and I try this with 1-14 simultaneous threads. The plot below is the speedup factor. Its pretty impressively linear, up to about 9 threads, then things get slower but then recover again. Not sure if that's meaningful - I need to run this 100 times and average.
14 March 2009
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:
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.