Profiling on EC2

Choosing Instances:

• Create 2 Ubuntu 20.04 server instances on c5.9xlarge boxes
  • Specify CPU options such that you use 1 thread per core
  • Up storage to 60 GB
  • Keep General Purpose SSD
• Name one instance Terrier and name one instance OLTPBench
• Update the security group on the Terrier instance
  • Add inbound rule allowing for custom TCP protocol on port 15721

OLTPBench Machine Setup:

• Update the box and install packages

  • sudo apt update
    sudo apt-get install ant 
    sudo apt-get install postgresql-client-common
    sudo apt-get install postgresql-client-12

• Clone Matt's fork of oltpbench:

  •  git clone https://github.com/mbutrovich/oltpbench.git

• Checkout the noisepage branch:

  •  git checkout noisepage

• Run setup scripts:

  •  ant bootstrap
     ant resolve
     ant build

• Update the noisepage TPCC config file:

  • vim config/noisepage_tpcc_config.xml

  • Change localhost to IP address of Terrier machine in DBUrl
  • Choose scalefactor, loaderThreads, and terminals
  • Add <warmup>10</warmup> in work to start caching
  • Choose default weights or change to 100,0,0,0,0 if you want just NewOrder transactions

Terrier Machine Setup:

• Update the box and clone Terrier

  • sudo apt update
    git clone https://github.com/cmu-db/terrier.git

• Update Ubuntu version in script/installation/packages.sh
  • Change value of $VERSION check from 18.04 to 20.04
• Run packages.sh

  • sudo bash script/installation/packages.sh

• Choose query mode between Interpret and Compiled
  • If you want Compiled mode change change the line
    exec_query->Run(common::ManagedPointer(exec_ctx), execution::vm::ExecutionMode::Interpret);
    to
    exec_query->Run(common::ManagedPointer(exec_ctx), execution::vm::ExecutionMode::Compiled);
    in
    src/traffic_cop/traffic_cop.cpp
• Build Terrier in Release mode with Debug symbols

  • mkdir RelWithDebug
    cd RelWithDebug
    cmake .. -DCMAKE_BUILD_TYPE=RelWithDebInfo -DTERRIER_USE_ASAN=OFF -DTERRIER_USE_JEMALLOC=ON
    make -j18 terrier tpl

• If you wish to log to ramdisk, set up ramdisk for WAL file storage

  • sudo mkdir /mnt/ramdisk
    sudo mount -t tmpfs -o rw,size=32G tmpfs /mnt/ramdisk 

• If you wish to use bcc profiling tools follow the installation steps described at the link

Running Scripts:

Terrier Machine:

• Note: you may want to use a tmux session in case you get disconnected from the instance
• Startup Terrier (if you do not want to log to Ramdisk, remove the log_file_path flag below)

  • sudo rm /mnt/ramdisk/wal.log 
    ./terrier -connection_thread_count=200 -wal_file_path=/mnt/ramdisk/wal.log

• Get debug symbols from used packages in Terrier using the RUNNING_PID_OF_TERRIER

  • echo "deb http://ddebs.ubuntu.com $(lsb_release -cs) main restricted universe multiverse
    deb http://ddebs.ubuntu.com $(lsb_release -cs)-updates main restricted universe multiverse
    deb http://ddebs.ubuntu.com $(lsb_release -cs)-proposed main restricted universe multiverse" | \
    sudo tee -a /etc/apt/sources.list.d/ddebs.list
    sudo apt install ubuntu-dbgsym-keyring
    sudo apt-key adv --keyserver keyserver.ubuntu.com --recv-keys F2EDC64DC5AEE1F6B9C621F0C8CAB6595FDFF622
    sudo apt-get update
    sudo apt install debian-goodies
    find-dbgsym-packages $(pidof terrier)

• Get debug symbols from remaining packages by taking last line of OUTPUT from previous command and doing:

  • sudo apt-get install **[OUTPUT]**

OLTPBench Machine:

• Load tables

  • ant execute -Dconfig=./config/noisepage_tpcc_config.xml -Dbenchmark=tpcc -Dexecute=false -Dload=true -Dcreate=true -Dextra="-histograms"

Terrier Machine:

• Do one of the following (NOTE: perf is the preferred method as it provides accurate symbols)
  • Attach perf profiler: /usr/bin/perf record --freq=997 --call-graph dwarf -q -p `pgrep terrier`
  • Attach bcc profiler: sudo profile-bpfcc -adf -F 99 -p `pgrep terrier` > ~/tpcc.profile
  • Profile offcputime with bcc: sudo offcputime-bpfcc -df -p `pgrep terrier` > ~/tpcc.offcpu

OLTPBench Machine:

• Run workload

  • ant execute -Dconfig=./config/noisepage_tpcc_config.xml -Dbenchmark=tpcc -Dexecute=true -Dload=false -Dcreate=false -Dextra="-histograms"

Visualizing Output:

Choose a path based on the tool used

PERF:

• In the directory where you recorded the profiling results on the Terrier Machine, produce a perf report
  • perf script > out.perf
• Clone the FlameGraph git repo and navigate to the directory
  • git clone https://github.com/brendangregg/FlameGraph.git
• Copy the output file to the FlameGraph directory and navigate to it
  • sudo cp out.perf ~/FlameGraph/
  • cd ~/FlameGraph
• Fold the stack samples into single lines
  • sudo ./stackcollapse-perf.pl out.perf > out.folded
• Produce a FlameGraph SVG
  • ./flamegraph.pl out.folded > kernel.svg
• Copy the SVG to your local machine and open it in a browser window

BCC:

• Copy over the file produced from your command above to your local machine
• Open speedscope.app and choose the output file

Important Notes:

• If you stop and restart the Terrier instance, you will have to delete and recreate the ramdisk storage you created:

  • sudo rm -rf /mnt/ramdisk
    sudo mkdir /mnt/ramdisk
    sudo mount -t tmpfs -o rw,size=32G tmpfs /mnt/ramdisk 

• Make sure to at least stop your instances after you are done profiling to avoid wasting credits