README.md

Snoopy: A Scalable Oblivious Storage System

Snoopy is a high-throughput oblivious storage system that scales similarly to a plaintext storage system. This implementation contains our Snoopy system as described in the forthcoming SOSP’21 paper.

WARNING: This is an academic proof-of-concept prototype and has not received careful code review. This implementation is NOT ready for production use.

This prototype is released under the Apache v2 license (see License).

Artifact Evaluation

Overview

• Baselines
  • Oblix (5 compute-minutes)
  • Obladi (15 compute-minutes)
  • Propagate Baselines (<1 minute)
• Provisioning Machines for Snoopy (5 compute-minutes)
• Figure 9 (195 compute-minutes)
• Figure 10a (35 compute-minutes)
• Figure 10b (10 compute-minutes)
• Figure 11a, 11b, 11c (6 compute-minutes)
• Figure 12a (2 compute-minutes)
• Figure 12b (12 compute-minutes)
• Figure 13 (160 compute-minutes)
• Figure 14a, 14b (2 compute-minutes)
• Cleanup (10 compute-minutes)
• Comparing Figures (20 human-minutes)

For our experiments, we will use a cluster of Azure confidential computing instances. Reviewers should have been provided with credentials to our Azure environment. Due to Azure resource quotas, we can only support one reviewer evaluating at a time. The entire artifact evaluation takes around 30 human-minutes and 7 compute-hours to complete.

All experiments produce .dat files in subfolders of ~/snoopy/experiments/results. These files are used to plot each figure that appears in the paper's evaluation, which are output in ~/snoopy/experiments/artifact_figures. All of the data files used in the paper's evaluation graphs can be found in ~/snoopy/experiments/artifact_figures/paper.

Only the baselines can be run concurrently. Both baselines must finish running before you start running any of the figures. The figures can be run in any order, except Figure 14, which must be run after Figure 11a, 11b, and 11c.

Each figure script is pretty automatic, i.e., you enter one command and the experiment(s) for that figure are run and plotted for you. Each experiment maintains a progress bar to give you an estimate of how much time is remaining.

Baselines

The Oblix and Obladi baselines can be run in parallel because they use different clusters. We recommend running the commands in two tmux windows.

Oblix

cd ~/snoopy/scripts/oblix
python3 runExperiment.py --config=config/oblix_benchmark.json -psrc     # 5 minutes

The above command can be broken down as follows:

• -p launches a cluster of 1 Azure SGX VM with an image that has Oblix installed.
• -s sets up a new experiment directory and symlinks it to ~/snoopy/experiments/oblix_benchmark/latest.
• -r runs the Oblix benchmark, which measures the latencies of accessing oblivious storages containing 10-2M blocks. We use the latency of accessing 2M blocks in Figure 9 (in order to calculate the baseline throughput), Figure 10b (baseline latency), Figure 13 (in order to emulate using Oblix as a subORAM).
• -c destroys the cluster.

Obladi

cd ~/snoopy/scripts/obladi
python3 runExperiment.py -psrc      # 15 minutes

The above command can be broken down as follows:

• -p launches an Obladi cluster of 2 VMs.
• -s sets up a new experiment directory and symlinks it to ~/snoopy/experiments/diff-clients-base-oram-par-server-1/latest. It compiles Obladi and sends the compiled jar files to the remote Obladi cluster.
• -r runs the Obladi benchmark, which measures the throughput of accessing an oblivious storage containing 2M blocks. We use this throughput to plot Figure 9.
• -c destroys the cluster.

Propagate Baselines

After running the Obladi and Oblix baselines, run the following command to propagate their values for future figures and experiments:

cd ~/snoopy/scripts
./process_baselines.sh      # <1 minute

Provisioning Machines for Snoopy

cd ~/snoopy/scripts
./provision_machines.sh     # 5 minutes

The above command launches a cluster of 21 Azure SGX VMs and updates the experiment config files with the machines' IP addresses. It costs ~$370/per day to run this cluster, so please destroy the cluster when you are done with artifact evaluation.

The rest of the instructions assume that your current working directory is ~/snoopy/scripts.

Figure 9

./figure9.sh      # 195 minutes

This command runs 3 experiments and then plots the data for Figure 9 showing the affect of increasing machines on total system throughput.

The graph will be located in ~/snoopy/experiments/artifact_figures/9.pdf. The artifact evaluation graph is drawn on the left, and the graph in the paper is drawn on the right.

Sample Figure 9 output

There may be some variance between the artifact and paper graph, but their trends should strongly resemble each other. This is due to the artifact only measuring requests for 10 seconds/iteration instead of 20 seconds to save on time.

Figure 10a

python3 runExperiment.py -f 10a -srg        # 35 minutes

This command runs 1 experiment and then plots the data for Figure 10a showing how the number of total blocks in Snoopy increases with increasing the number of subORAMs while maintaining a mean request latency of 160ms.

The graph will be located in ~/snoopy/experiments/artifact_figures/10a.pdf. The artifact evaluation graph is drawn on the left, and the graph in the paper is drawn on the right.

Sample Figure 10a output

Again, there may be some variance between the artifact and paper graph, but their trends should strongly resemble each other. This is due to the artifact only measuring requests for 10 seconds/iteration instead of 20 seconds to save on time.

Figure 10b

python3 runExperiment.py -f 10b -srg        # 10 minutes

This command runs 1 experiment and then plots the data for Figure 10b showing how the mean request latency in Snoopy decreases with increasing the number of subORAMs.

The graph will be located in ~/snoopy/experiments/artifact_figures/10b.pdf. The artifact evaluation graph is drawn on the left, and the graph in the paper is drawn on the right.

Sample Figure 10b output

Figure 11a, 11b, 11c

./figure11.sh     # 6 minutes

This command runs 3 experiments and then plots the data for Figure 11a, 11b, and 11c, showing the performance breakdowns for processing a batch in Snoopy while varying the number of requests in each batch. Figure 11a shows this breakdown for a data size of 2^10 objects, Figure 11b shows 2^15 objects, and Figure 11c shows 2^20 total objects.

Sample Figure 11a, 11b, 11c output

Figure 12a

python3 runExperiment.py -f 12a -srg      # 2 minutes

This command runs 1 experiment and then plots the data for Figure 12a, showing the performance of bitonic sort using multiple threads.

Sample Figure 12a output

Figure 12b

python3 runExperiment.py -f 12b -srg      # 12 minutes

This command runs 1 experiment and then plots the data for Figure 12a, showing the performance of a subORAM processing a batch using multiple threads.

Sample Figure 12b output

Figure 13

./figure13.sh     # 160 minutes

This command runs 3 experiments and then plots the data for Figure 13 showing the affect of increasing machines on Snoopy where the subORAMs are replaced with Oblix.

Sample Figure 13 output

Again, there may be some variance between the artifact and paper graph, but their trends should strongly resemble each other. This is due to the artifact only measuring requests for 10 seconds/iteration instead of 20 seconds to save on time.

Figure 14a, 14b

PREREQUISITE: Requires Figure 11a, 11b, and 11c.

./figure14.sh     # 2 minutes

This command runs the planner using the microbenchmarks collected in Figure 11's experiments and plots Figures 14a, 14b which show the optimal system configuration as throughput requirements increase for different data sizes.

Sample Figure 14a, 14b output

Cleanup

python3 runExperiment.py -c     # 10 minutes

This command destroys the cluster. Please remember to run this when you are done with artifact evaluation. It costs ~$370/per day to run the cluster!

Comparing Figures

All figures produced by this artifact are located in ~/snoopy/experiments/artifact_figures. We recommend scp'ing this folder to your local computer so you can easily view each PDF. In each PDF, the artifact evaluation graph is drawn on the left and the graph in the paper is drawn on the right.

Sample Figure 9 output

Sample Figure 10a output

Sample Figure 10b output

Sample Figure 11a, 11b, 11c output

Sample Figure 12a output

Sample Figure 12b output

Sample Figure 13 output

Sample Figure 14a, 14b output

Playing with Snoopy

Building Snoopy Locally

Protobuf:

# On ubuntu 18.04
sudo apt-get install build-essential autoconf libtool pkg-config automake zlib1g-dev
cd third_party
git submodule update --init
cd protobuf/cmake
mkdir build
cd build
cmake -Dprotobuf_BUILD_TESTS=OFF -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=`pwd`/../install ..
make -j `nproc`
make install

gRPC (requires protobuf first):

cd grpc
git submodule update --init
mkdir build
cd build
cmake -DCMAKE_PREFIX_PATH=`pwd`/../../protobuf/cmake/install -DgRPC_INSTALL=ON -DgRPC_BUILD_TESTS=OFF \
      -DgRPC_PROTOBUF_PROVIDER=package -DgRPC_ZLIB_PROVIDER=package -DgRPC_CARES_PROVIDER=module -DgRPC_SSL_PROVIDER=package \
      -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_INSTALL_PREFIX=`pwd`/install \
      ../
make
make install

Need to put boost library in /usr/local/ (doesn't need to be built or installed).

Running anything in the scripts directory:

Follow instructions here: https://docs.microsoft.com/en-us/azure/developer/python/configure-local-development-environment?tabs=cmd

pip install -r requirements.txt
export AZURE_SUBSCRIPTION_ID=<azure-subscription-id>

Debugging

• Make sure DEBUG is turned on in enc.conf and OE_FLAG_DEBUG is passed into enclave creation

• Make debug build:

mkdir debug
cd debug
cmake -DCMAKE_BUILD_TYPE=Debug ..
make

Run oegdb as usual:

/opt/openenclave/bin/oegdb --args ...

Handle SIGILL:

(gdb) handle SIGILL nostop