[Discuss-gnuradio] multiprocessor flow graphs

[Dev Ramudit]

Hi all,

	I'm trying to modify the benchmark_ofdm_rx.py example to handle a 
decimation of 8 for a project I am working on. Going straight for the 
usrp->file works fine, so if I can handle the processing requirement I 
should be able to demod ofdm packets in real time. The PC I'm working 
with has 4 processors, but as far as I know, the gnuradio flow graphs 
should only be using one per program.  Therefore, I took the blocks from 
the various files used in benchmark_ofdm_rx and put them into a single 
long flow graph. I then split the flow graph and am trying to do the 
following:

if os.fork() == 0: # child process
        {first half of benchmark_ofdm_rx.py flow graph}
        topblock.connect(middleblock, shared_memory_sink)
        topblock.run()
else:   # parent process
        topblock.connect(shared_memory_source, nextblock)
        {second half of benchmark_ofdm_rx.py flow graph}
        topblock.run()

	Originally I tried this with the shared_memory_source/sink being the 
udp_source/sink blocks provided with gnuradio. The modified flow graph 
worked perfectly at high decimations, and seemed to be utilizing two 
processors. However, as I got closer to 8 I would "lose" packets across 
the udp source/sink. This happened no matter what I set the MTU to. I 
tried using fifos/pipes instead, but there is apparently a 64K memory 
limit in the linux kernel. My current idea is to use the circular buffer 
classes that Gnuradio has modified to use non-private shared memory 
source/sinks. From what I've understood from Gnuradio these circular 
buffers are what link up the blocks in a graph, so it should be fine for 
handling the interprocess throughput I need. Is the best way to 
accomplish what I'm trying to do?

Thanks,
Dev