Skip to content

Latest commit

 

History

History
59 lines (40 loc) · 3.28 KB

README.md

File metadata and controls

59 lines (40 loc) · 3.28 KB

Cloud Deployment

This folder contains scripts and instructions for reproducing the experiments conducted under geo-distributed settings in our arXiv '21 paper. Our training evaluations rely on a distributed setting of multiple machines under the Parameter-Server (PS) architecture. In our paper, we used 10 c5.4xlarge machines to simulate 10 participants and 1 r5.4xlarge machine to emulate the server in each round. To be exact, their configurations are:

  • OS kernel image: ubuntu-bionic-18.04-amd64-server-20200903;
  • Computational capabilities: 16 vCPUs, 32 GB (128 GB) memory (with 16G swap) and 100 GiB SSD Storage.

Overview

Prerequisites

In order to have FATE v1.2.0 run at your cluster, in addition to using similar computational capabilities as what we played with, the network connectivity also needs to meet the following requirements:

  1. Nodes within should be able to reach one another via private IPs.
  2. Nodes should be able to perform remote login to one another without password or identification files specified.

If you are willing to renting resources offered by AWS like we did, we are happy to provide you with some handy scripts at ../utils, which should help you prepare an eligible cluster from scratch. Check out the README there for details.

Cluster Setup

Once you have successfully established a several-node cluster which meets the two aforementioned prerequisites, be relaxed as all the necessary setup commands have been packed into one script or two. Note that some commands related to FATE v1.2.0's deployment may be intrusive to your systems. For example, it will extend some system limits and install dependencies in a global sense. If you have concerns about these, you may want to delve into the script first and see if you need to switch to another evaluation environment.

  1. At all nodes (server & clients), execute the following command. 
    cd /home/ubuntu/FLASHE/deployment
source ./all_nodes_prepare.sh
  2. Then at server node, make appropriate modifications on the configuration file cluster_conf.yml (comments there can help you quickly understand how to specify the parameters) and run 
    source ./server_deploy.sh cluster_conf.yml

Job Configuration

Evaluation

NOTE: .

The output of the experiment will validate the following major claims in our paper:

  1. FLASHE outperforms batching versions of the three baselines by 3.2×-15.1× in iteration time and 2.1×-42.4× in network footprint. Compared to plaintext training, FLASHE achieves near-optimality efficiency with overhead ≤6% and 0% in traffic (§6.2) —> Figure 7 and 8.
  2. FLASHE brings down at least 13×-63× of the computation overhead and 48× of the communication overhead of general HEs when sparsification is employed (§6.4) —> Figure 10 and 11.
  3. FLASHE's mask precomputation effectively eliminates the cryptographic overhead (rendering it < 0.1s) (§6.6) —> Figure 13.

Per Iteration Efficiency (Figure 7 and Figure 8)

Figure 7

Figure 8