clearlinux/graphene
2
0
Fork 0
mirror of https://github.com/clearlinux/graphene.git synced 2026-09-05 05:12:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

FOR TESTING ONLY: Test with different batch sizes, fix brk size

Browse Source
This commit is contained in:
Dmitrii Kuvaiskii
2020-08-21 14:16:44 -07:00
committed by root
parent 3422cf1cb6
commit e9f78660d8
4 changed files with 57 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Examples/pytorch/pytorch.manifest.template.ppml
+11 -2
View File
@@ -19,6 +19,13 @@ loader.debug_type = $(GRAPHENEDEBUG)
# Read application arguments directly from the command line. Don't use this on production!
loader.insecure__use_cmdline_argv = 1
# PyTorch may use up to 28GB of RAM in our experiments, so disable ASLR to have more enclave memory
loader.insecure__disable_aslr = 1
# PyTorch/Python uses up to 10GB of brk region in our experiments, so set this option. Otherwise,
# expect significant perf degradation since PyTorch falls back to slower mmap/munmap logic.
sys.brk.max_size = 12G
# Environment variables
loader.env.LD_LIBRARY_PATH = /lib:/usr/lib:$(ARCH_LIBDIR):/usr/$(ARCH_LIBDIR):./
loader.env.LD_PRELOAD = "libsecret_prov_attest.so /lib/libgomp.so.1"
@@ -79,7 +86,7 @@ fs.mount.pip.uri = file:$(HOME)/.local/lib
# than the enclave size, Graphene will not be able to allocate it.
#
# In particular, libtorch*.so is more than 1G, thus 4G is the minimum to make this run.
sgx.enclave_size = 8G
sgx.enclave_size = 32G
# Set the maximum number of enclave threads. For SGX v1, the number of enclave
# TCSes must be specified during signing, so the application cannot use more
@@ -89,7 +96,9 @@ sgx.enclave_size = 8G
# the application can create is (sgx.thread_num - 2).
#
# We (somewhat arbitrarily) specify 128 threads for this workload.
sgx.thread_num = 128
sgx.thread_num = 176
#sgx.rpc_thread_num = 176
# SGX trusted libraries
Examples/pytorch/pytorch.manifest.template.sgx
+10 -1
View File
@@ -19,6 +19,13 @@ loader.debug_type = $(GRAPHENEDEBUG)
# Read application arguments directly from the command line. Don't use this on production!
loader.insecure__use_cmdline_argv = 1
# PyTorch may use up to 28GB of RAM in our experiments, so disable ASLR to have more enclave memory
loader.insecure__disable_aslr = 1
# PyTorch/Python uses up to 10GB of brk region in our experiments, so set this option. Otherwise,
# expect significant perf degradation since PyTorch falls back to slower mmap/munmap logic.
sys.brk.max_size = 12G
# Environment variables
loader.env.LD_LIBRARY_PATH = /lib:/usr/lib:$(ARCH_LIBDIR):/usr/$(ARCH_LIBDIR)
loader.env.LD_PRELOAD = "/lib/libgomp.so.1"
@@ -65,7 +72,7 @@ fs.mount.pip.uri = file:$(HOME)/.local/lib
# than the enclave size, Graphene will not be able to allocate it.
#
# In particular, libtorch*.so is more than 1G, thus 4G is the minimum to make this run.
sgx.enclave_size = 8G
sgx.enclave_size = 32G
# Set the maximum number of enclave threads. For SGX v1, the number of enclave
# TCSes must be specified during signing, so the application cannot use more
@@ -77,6 +84,8 @@ sgx.enclave_size = 8G
# We (somewhat arbitrarily) specify 128 threads for this workload.
sgx.thread_num = 176
#sgx.rpc_thread_num = 176
# SGX trusted libraries
sgx.trusted_files.ld = file:$(GRAPHENEDIR)/Runtime/ld-linux-x86-64.so.2
Examples/pytorch/pytorchexample.py
+10 -18
View File
@@ -1,21 +1,21 @@
# This PyTorch image classification example is based off
# https://www.learnopencv.com/pytorch-for-beginners-image-classification-using-pre-trained-models/
from timeit import default_timer as timer
from torchvision import models
import statistics
import torch
from torchvision import models
torch.set_num_threads(72)
EXP_NUM_THREADS = 1
EXP_BATCH_SIZE = 1
torch.set_num_threads(EXP_NUM_THREADS)
# Load the model from a file
start = timer()
#alexnet = torch.load("pretrained.pt")
alexnet = models.inception_v3(pretrained=True)
end = timer()
model_load_time = end - start
# Prepare a transform to get the input image into a format (e.g., x,y dimensions) the classifier
# expects.
from torchvision import transforms
transform = transforms.Compose([
transforms.Resize(256),
@@ -26,23 +26,20 @@ transform = transforms.Compose([
std=[0.229, 0.224, 0.225]
)])
# Load the image.
from PIL import Image
img = Image.open("input.jpg")
# Apply the transform to the image.
img_t = transform(img)
# Magic (not sure what this does).
batch_t = torch.unsqueeze(img_t, 0)
batch_t = torch.unsqueeze(img_t, 0).repeat(EXP_BATCH_SIZE, 1, 1, 1)
alexnet.eval()
for i in range (0,3):
# warmup runs
for i in range (0,10):
out = alexnet(batch_t)
# experiment runs
sample = []
for i in range (0,200):
for i in range (0,30):
start = timer()
out = alexnet(batch_t)
end = timer()
@@ -50,16 +47,11 @@ for i in range (0,200):
print(model_load_time, statistics.mean(sample),statistics.stdev(sample))
# Load the classes from disk.
with open('classes.txt') as f:
classes = [line.strip() for line in f.readlines()]
# Sort the predictions.
_, indices = torch.sort(out, descending=True)
# Convert into percentages.
percentage = torch.nn.functional.softmax(out, dim=1)[0] * 100
# Print the 5 most likely predictions.
with open("result.txt", "w") as outfile:
outfile.write(str([(classes[idx], percentage[idx].item()) for idx in indices[0][:5]]))
Examples/pytorch/run-experiment.sh
+26 -31
View File
@@ -1,6 +1,6 @@
#!/bin/bash
# BEFORE RUNNING THIS:
# BEFORE RUNNING THIS, FOR SGX PPML:
# - Build Examples/ra-tls-secret-prov with DCAP and start ./secret_prov_server_dcap
# - Build libsecret_prov_attest.so and copy here
@@ -9,6 +9,15 @@ set -e
SCRIPT=pytorchexample.py
STDERRFWD=stderr.log
# batch size of 1 best for latency, batch size of 32 best for throughput;
# we use 32 because higher batch sizes (64, 128) lead to 64GB enclaves and
# are thus too slow for our experiments
sizes=(
1
32
)
# experiments performed on 36 physical cores, 72 hyperthreads
threads=(
1
2
@@ -25,12 +34,13 @@ threads=(
72
)
# for experiments, use only 4 most interesting network models
networks=(
# alexnet
squeezenet1_0
vgg19
wide_resnet50_2
resnet50
# alexnet
# densenet161
# mobilenet_v2
# googlenet
@@ -57,57 +67,41 @@ else
fi
for i in {1..3}; do
for net in "${networks[@]}"; do
for SIZE in "${sizes[@]}"; do
for NETWORK in "${networks[@]}"; do
for NUM_THREADS in ${threads[@]}; do
echo "=== $1 $net $i $NUM_THREADS ==="
echo "=== $1 $i $SIZE $NETWORK $NUM_THREADS ==="
# set the number of threads
sed "s/set_num_threads([0-9]*)/set_num_threads($NUM_THREADS)/" -i $SCRIPT
# set the number of threads, batch size, and used model
sed "s/EXP_NUM_THREADS = [0-9]*/EXP_NUM_THREADS = $NUM_THREADS/" -i $SCRIPT
sed "s/EXP_BATCH_SIZE = [0-9]*/EXP_BATCH_SIZE = $SIZE/" -i $SCRIPT
sed "s/models.[a-zA-Z0-9_]*(/models.$NETWORK(/" -i $SCRIPT
# change the python script to switch the model
sed "s/models.[a-zA-Z0-9_]*(/models.$net(/" -i $SCRIPT
###########################
# Native #
###########################
if [ "$1" == "native" ]
then
if [ "$1" == "native" ]; then
make clean >/dev/null
make SGX=1 >/dev/null
OMP_NUM_THREADS=$NUM_THREADS MKL_NUM_THREADS=$NUM_THREADS numactl --cpunodebind=0 --membind=0 \
python3 $SCRIPT
fi
###########################
# Graphene w/o SGX #
###########################
if [ "$1" == "graphene" ]
then
if [ "$1" == "graphene" ]; then
make clean >/dev/null
make SGX=1 >/dev/null
OMP_NUM_THREADS=$NUM_THREADS MKL_NUM_THREADS=$NUM_THREADS numactl --cpunodebind=0 --membind=0 \
./pal_loader pytorch.manifest $SCRIPT 2>$STDERRFWD
fi
###########################
# Graphene w/ SGX #
###########################
if [ "$1" == "sgx" ]
then
if [ "$1" == "sgx" ]; then
make clean >/dev/null
NETFILE=$(ls ~/.cache/torch/hub/checkpoints | grep "^$net")
NETFILE=$(ls ~/.cache/torch/hub/checkpoints | grep "^$NETWORK")
sed "s/[a-z0-9_-]*\.pth$/$NETFILE/" -i pytorch.manifest.template
make SGX=1 >/dev/null
SGX=1 OMP_NUM_THREADS=$NUM_THREADS MKL_NUM_THREADS=$NUM_THREADS numactl --cpunodebind=0 --membind=0 \
./pal_loader pytorch.manifest $SCRIPT 2>$STDERRFWD
fi
###########################
# Graphene w/ SGX + PPML #
###########################
if [ "$1" == "ppml" ]
then
sed "s/models\.[a-zA-Z0-9_]*(/models\.$net(/" -i download-pretrained-model.py
if [ "$1" == "ppml" ]; then
sed "s/models\.[a-zA-Z0-9_]*(/models\.$NETWORK(/" -i download-pretrained-model.py
make clean >/dev/null
make download_model >/dev/null
LD_LIBRARY_PATH=../ra-tls-secret-prov ../ra-tls-secret-prov/pf_crypt encrypt -w ../ra-tls-secret-prov/files/wrap-key -i ./plaintext/pretrained.pt -o pretrained.pt >/dev/null
@@ -120,3 +114,4 @@ for NUM_THREADS in ${threads[@]}; do
done
done
done
done