Hello everyone,
I'm seeing a difference in performance when running a CPU intensive pipeline using the command-line tool
gpt and from inside Java code.
With
gpt, graph execution takes about 9 minutes and
htop shows all 8 CPU cores at 100% utilization.
With the Java code, the same graph takes 15 minutes. According to
htop, on average 2 CPU cores are being used at 100% (average of 25% CPU usage).
I've tried setting the system properties
beam.parallelism and
seadas.parallelism with no noticeable improvement.
I'm invoking the
GPF from Java using the following code:
1
2ProgressMonitor pm;
3...
4Graph graph = GraphIO.read(new StringReader(graphStringI));
5GraphProcessor processor = new GraphProcessor();
6processor.executeGraph(graph, pm);
Some further information which might be useful:
- I'm running Ubuntu 12.04 on a 64-bit machine with 8GB RAM and a Corei7-3770 CPU.
- Using SeaDAS 7.0.2 and Open JDK (1.7.0_55).
- The graph XML produces 7 day averages of MODIS aqua chlorophyll data over the Southern GBR region for a month.
- The pipeline has been made artificially CPU intensive, by performing extra unrelated calculations inside computeTile() of the band averaging operator.
- The Java code is run as a junit test case from Maven.
(My question is fairly similar to
this forum question, but with some differences as mentioned above.)
What could I be missing here?
thanks,
Kutila
Answer