SmartDIMM: In-Memory Acceleration of Upper Layer I/O Protocols Artifact

This repository contains scripts for SmartDIMM HPCA'24 artifact evaluation of the SmartDIMM: In-Memory Acceleration of Upper Layer I/O Protocols Artifact paper by Neel Patel, Amin Mamandipoor, and Mohammad Alian

HPCA 2024 Artifact Evaluators should skip to Artifact Evaluation instructions below

Evaluation instructions

• start by executing git submodule update --init --recursive to fetch all submodules and dependencies The scripts in this repository can be used to reproduce:

• Figure 10: SmartDIMM Scratchpad Utilization

• Figure 11: Encrypted Nginx Performance (Nginx Server for Encrypted Files) Measures RPS, CPU Utilization, and Memory Bandwidth for a server producing TLS-encrypted HTTP responses driven by the wrk workload generator

• Figure 12: Compressed Nginx Performance (Nginx Server for Compressed Files) Measures RPS, CPU Utilization, and Memory Bandwidth for a server producing gzip-compressed HTTP responses driven by the wrk workload generator

• As resource availability of nodes on the Clemson cluster equipped with Mellanox Bluefield 2 SmartNICs cannot be guaranteed we provide separate instructions for our artifact evaluators below.

Directory Structure

|--- Near-Memory-Sensitivity-Analysis: SmartDIMM sensitivity analysis examining the resource utilization for different server loads
|--- wrk_offloadenginesupport: workload generation scripts based on the `wrk` http request generation tool
  |--- async_nginx_build: nginx server configuration files and builds for baseline http, https, and accelerated https using QAT and kTLS

Setting up Cloudlab Instances:

To ease reproduction we have created a cloudlab environment and setup closely matching the server configuration used in the HPCA 2024 paper SmartDIMM: In-Memory Acceleration of Upper Layer I/O Protocols Artifact

Following the instructions below will provision a cloudlab instance in which the SPEC 2017 and (de)compression workloads from the paper will be executed. For more information, refer to [SmartDIMM: In-Memory Acceleration of Upper Layer I/O Protocols Artifact](https://www.HPCA 56.org/)

• allocate a cloudlab instance using the genilib script provided in this repo
  • Create a cloudlab account if needed
  • Navigate to Experiments, then Create Experiment Profile, and upload nginx_workload.profile

Compiling DUT servers

• on DUT node

git clone git@github.com:architecture-research-group/SmartDIMM_ArtifactEvaluation.git
git submodule update --init --recursive
cd wrk_offloadenginesupport/async_nginx_build
make all

Compiling Workload Generators

• on Workload Generator node

git clone git@github.com:architecture-research-group/SmartDIMM_ArtifactEvaluation.git
git submodule update --init --recursive
cd wrk_offloadenginesupport
make all

Artifact Evaluation Instructions:

1. Install Cisco AnyConnect VPN client. You can use the following link to request access to download the client for Mac, windows, or Linux. We will give you access immediately after your request.
2. Run Cisco AnyConnect client.
3. Contact SmartDIMM authors over hotcrp site to ask for a temporary password so you use it in the next step. We will do our best to make ourselves available at the reviewer's request.
4. Enter kuanywhere.ku.edu for the address, and hit "connect".

5. Select Group as "DuoAuthentication" Username: sgupta_sta Password: hpca2024_AE Second Password: [The temporary password that Siddharth Gupta Provides]

- Alternatively: type "push" into the Second Password field and notify Siddharth Gupta 6. Now you are connected to the VPN and should be able to SSH to the servers.

Logging in to the servers can be performed over ssh once connected to the VPN:

ssh sgupta@castor.ittc.ku.edu
ssh sgupta@pollux.ittc.ku.edu
ssh sgupta@axdimm.ittc.ku.edu

• above commands provide access to the servers used throughout the subsequent steps

Nginx Workload Experiments

• Corresponds to figures 11 and 12 in SmartDIMM: In-Memory Acceleration of Upper Layer I/O Protocols Artifact
  

ssh sgupta@castor.ittc.ku.edu #login to workload generator
cd /home/n869p538/wrk_offloadenginesupport

./run_tls.sh # performs max_rps tls test from fig. 11
			 # prints results to stdout (Normalize to http configuration for RPS comparison in Fig. 11)
./run_tls_const.sh # performs system resource utilization tls test form fig.11
			 # prints results to stdout (Normalize to http configuration for RPS comparison in Fig. 11)

./run_gzip.sh # performs max_rps nginx gzip server test from fig.12
			  # prints results to stdout (Normalize to http configuration for RPS comparison in Fig. 11)
./run_gzip_const.sh # performs system resource utilization nginx gzip server test from fig.12
			  # prints results to stdout (Normalize to http configuration for RPS comparison in Fig. 12)

SmartDIMM Sensitivity Analysis

• Corresponds to figure 10 in SmartDIMM: In-Memory Acceleration of Upper Layer I/O Protocols Artifact
  

FPGA Programming (Control PC)

• To program the bit file onto the FPGA of the AXDIMM and activate the ILAs for result collection, follow these steps:

# Access pollux (Control PC) via SSH
ssh sgupta@pollux.ittc.ku.edu             

# Navigate to the sensitivityAnalysis directory
cd /home/sgupta/sensitivityAnalysis/vivadoScript

# Execute the run.sh script with specified FolderName and FileName
source /tools/xilinx/Vivado/2019.2/settings64.sh
./run.sh FolderName FileName 

Starting Workload Generation:

• Initiates ComputeCopy Threads performing memcpy's to/from the AxDIMM's registered address space

# Access axdimm via SSH
ssh sgupta@axdimm.ittc.ku.edu           

# Navigate to the test directory
cd /home/sgupta/axdimm_scripts/antagonist_testing

# Execute the antagonist.sh alongside test.sh script 
./antagonist.sh & ./test.sh

Generating Sensitivity Analysis Figures

• Lastly, parse the collected CSV files and generate plots using the following commands:

# Navigate to the `FolderName` directory specified in 
cd /home/sgupta/sensitivityAnalysis

python3 emulation.py -csv PathToCSVFile-0 PathToCSVFile-1 PathToCSVFile-3