Skip to content

Configurations

HPMPC provides extensive configuration options to customize MPC programs, optimize their performances, and support various hardware architectures. Understanding the configuration options is crucial to achieving the best performance for a specific use case. The following sections provide an overview of the most important configuration options and tips on how to set them. A complete list of configuration options can be found in the config.h file. By default, the config.h file contains tweaks that should work well for most use cases.

To change a configuration option, specify the option in the Makefile for temporary changes or in the config.h file for permanent changes that also apply to all future compilations. By default, the Makefile sets the configuration options to the ones in config.h. Every configuration option can be overridden by the Makefile by specifying the option as an argument to the make command. 

make -j PARTY=<party_id> FUNCTION_IDENTIFIER=<function_identifier> NUM_INPUTS=<num_inputs>

Basic Settings

Configuration Option Description Tips
PROTOCOL Protocol to use for the MPC computation. Protocol 5 (Trio) is recommended for 3PC, Protocol 12 (Quad) for 4PC.
PRE Use a preprocessing phase? Preprocessing is supported by Protocols 3, 5, 8, 11, 12.
PARTY Party ID (starting from 0). For local execution, PARTY=all can be set by the Makefile.
BITLENGTH Bitlength of integers. Lower BITLENGTHs can improve performance.
FRACTIONAL Fractional bits to use for fixed-point arithmetic. Higher values can improve precision but perform worse with probabilistic truncation.
FUNCTION_IDENTIFIER Identifier of the function to run. Refer to protocol_executer.hpp for a list of different predefined functions.
NUM_INPUTS Number of inputs. Used by benchmarking functions or neural networks to set the number of inputs processed (multiplies with concurrency settings and `SPLITROLES).

Concurrency Settings

Configuration Option Description Tips
DATTYPE Register size to use for SIMD parallelization. Use the maximum supported by the hardware (e.g. 512 if the CPU supports AVX-512). Note that some BITLENGTHs are not supported by all DATTYPEs.
PROCESS_NUM Number of parallel processes to use. Experiment with different values to find the optimal performance.

Hardware Acceleration Settings

Configuration Option Description Tips
RANDOM_ALGORITHM Random number generation algorithm. 2 uses VAES or AES-NI, and 1 uses Bitslicing. Prefer 2 for performance.
USE_SSL_AES Use OPENSSL's AES implementation. Prefer 0 for performance.
ARM ARM processor support. Set to 1 for ARM processors, 0 otherwise.
USE_CUDA_GEMM Use CUDA for matrix multiplication. Prefer 2 or 4 for performance.

Tweaks

Configuration Option Description Tips
SEND_BUFFER Number of messages to buffer before sending. Experiment with different values to find the optimal performance.
RECV_BUFFER Number of receiving messages to buffer. Experiment with different values to find the optimal performance.
VERIFY_BUFFER Number of messages to buffer before hashing. The lower the better, but must be at least 512/DATTYPE for correctness.

Neural Network Settings

Configuration Option Description Tips
MODELOWNER Who holds the model parameters? Use P_0 for party 0, P_1 for party 1, etc. Use -1 to use dummy model parameters.
DATAOWNER Who holds the data? Use P_0 for party 0, P_1 for party 1, etc. Use -1 to use dummy data.
TRUNC_THEN_MULT Truncate before or after multiplication. Set to 1 for higher accuracy in most cases.
TRUNC_APPROACH Truncation approach. Use 0 for probabilistic truncation, 1 for reduced slack truncation, 2 for exact truncation Prefer 0 for performance and 1 or 2 for accuracy.
TRUNC_DELAYED Delay CONV truncation until next ReLU. Must be set to 1 for truncation approaches 1 and 2.
COMPUTE_ARGMAX Compute final Argmax during inference. Set to 0 to reveal class probabilities.
PUBLIC_WEIGHTS Use public weights. Set to 1 to use public weights for improved performance.
COMPRESS Use 8-bit ReLUs for all layers. Set to 1 to benchmark ReLUs with reduced bitlength.
REDUCED_BITLENGTH_k cut off BITLENGTH-k least significant bits before sign bit extraction. May reduce accuracy significantly.
REDUCED_BITLENGTH_m cut off m most significant bits before sign bit extraction. May reduce accuracy significantly.
IS_TRAINING Training or inference phase. Training is not supported yet.

For different adder approaches there are also the ONLINE_OPTIMIZED and BANDWIDTH_OPTIMIZED options, where

  • BANDWIDTH_OPTIMIZED=1 uses a Ripple Carry Adder (RCA) approach.
  • BANDWIDTH_OPTIMIZED=0 and ONLINE_OPTIMIZED=0 uses a Parallel Prefix Adder (PPA) approach.
  • BANDWIDTH_OPTIMIZED=0 and ONLINE_OPTIMIZED=1 uses a 4-Way parallel prefix adder approach (PPA4).

For the predefined neural networks a FUNCTION_IDENTIFIER<100 uses RCA, FUNCTION_IDENTIFIER<200 uses PPA, and FUNCTION_IDENTIFIER<300 uses PPA4. For instance FUNCTION_IDENTIFIER=74 executes VGG16 with RCA, FUNCTION_IDENTIFIER=174 with PPA, and FUNCTION_IDENTIFIER=274 with PPA4.