Build instructions:

• build BMI shared libraries
• build CCI shared libraries (downloads at http://cci-forum.com/?page_id=46)
• build mercury shared libraries
• set PKG_CONFIG_PATH to point at mercury's pkgconfig path
• set LD_LIBRARY_PATH to point at BMI's, CCI's library paths
• run make with appropriate CFLAGS (e.g. -O3)

Benchmark descriptions

NOTE: hg-ctest4 is most likely the one you should use - tests 1-3 are looking at very specific threading scenarios.

hg-ctest1

• looking at client-side concurrent issuance of data, rpc's
• use implicit bulk context created by HG_Init, issue bulk push followed by RPC call to (optionally different) servers N times

hg-ctest2

• concurrent rpc/bulk ops in different threads
• uses different HG contexts backed by a single NA context
• benchmark runs a fixed amount of time, each thread issues back to back calls

hg-ctest3

• single-threaded version of hg-ctest2 (same HG context, same thread executing rpc / bulk calls in whatever order HG_Trigger hands callbacks out)

hg-ctest4

• N client processes, 1 server process, each single-threaded. Clients can be configured in RPC or bulk xfer mode.

Running

general

• when running the server, it's useful to fix the port the server attaches to. For that purpose, use CCI_CONFIG=cci-serv.conf (modify to your purposes) when running ./hg-ctestX server.
  • alternatively, the server spits out it's address to the file "ctest1-server-addr.tmp", so you can also script against that.
• run programs without arguments to see usage instructions

provided scripts

NOTE: you will likely need to lightly modify the scripts to use them successfully.

• run-ctest.sh - an ssh-based benchmark runner. A bit hairy, but should be able to do basic benchmarks.
• run-local-prof.sh - locally runs and optionally profiles the hg-ctest4 benchmark.
• runall-cooley.sh - performs a collection of benchmark runs on the ALCF "cooley" cluster (http://www.alcf.anl.gov/resources-expertise/analytics-visualization). Uses Cobalt environment variables. Has a lot of copy-paste at the moment :-/.

Scratch notes

• CCI ignores mercury URIs passed in, but Mercury must be able to correctly parse the URI input. So just use a dummy like cci+tcp://foobar as the URI for the server (the client of course needs to provide the resulting URI)
• As a complete hack, in certain cases running in single-threaded mode one can disable pthread usage. Specify USE_DUMMY_PTHREAD=yes in your make invocation.