This source code is released under a BSD-style license. See LICENSE
for more details.


I. Installation
  
  Our code can either use Open MPI or HTTP for communication between the
  client and server; in our experiments, we used Open MPI. The following
  are the dependent packages, if installing in the Open MPI mode:

  1. gmp with C++ extensions enabled
  2. PAPI
  3. Chacha pseudorandom number generator (use the assembly version for
  high performance)
  4. NTL with NTL_GMP_LIP=on 
  5. libconfig
  6. Cheetah template library
  7. OpenMPI
  8. CUDA packages (if USE_GPU is set to 1 in GNUMakefile)
  9. GMPMEE 

II. Configuration

  - GPU usage is disabled by default; to enable check the USE_GPU flag
    in code/pepper/flags.
   
  - By default, our build system selects the Zaatar protocol (see below
    for automatically selecting a different protocol); however, to
    manually select another protocol, modify the variable FRAMEWORK in
    code/pepper/flags (e.g. "FRAMEWORK=GINGER").

III. Running examples
  
  (1) With a single machine, for testing purposes

  Suppose the computation is apps_sfdl/mm_sfdl.sfdl, then do the
  following.

    Specify input sizes in apps_sfdl/mm_sfdl.sfdl

    Run "make SFDL_FILES=mm_sfdl"
    
    [If the computation is expressed in C, then supply the computation
    to compile via "C_FILES=<file_name>"]

    #Note - The first time a new computation is being compiled, the above command will
    #fail with "No target to compile ... _p_exo.o". In this case, re-run
    #the above command and it should work.
    
    To test the Ginger/Zaatar protocol, run "./run/mm_sfdl.sh"

      To run a computation using the CMT-batching protocol, build it
      with FRAMEWORK=ZAATAR, but then run with
      ./bin/cmtgkr/apps/pws_cmtgkr bin/[computation_name.pws]
      [batch-size] [num of bits in input's numerator] [num of bits in
      input's denominator]
  
  (2) With a cluster, for experiments.

    We run our experiments using the longhorn cluster where the client
    and the server run on different machines and the server can have
    access to multiple cores/machines. Our papers at Usenix Security
    2012, EuroSys 2013, and S&P (Oakland) 2013 contain details of the
    experimental configuration.

    Please refer to
    http://www.tacc.utexas.edu/user-services/user-guides/longhorn-user-guide
    for launching a job on TACC. To get started, we are releasing the
    scripts that we used to launch our experiments on TACC; they are
    under the "scripts" folder (scripts/run_zaatar_jobs.py and
    scripts/run_allspice_jobs.py), and the script
    scripts/allspice_beta.py allows one to select a protocol (among
    {Ginger, Zaatar, and CMT-batching}) with the lowest break-even batch
    sizes.

IV. Contact
Please contact srinath at cs dot utexas dot edu for any questions and
comments.