TriCache Artifact Evaluation

Hardware & Software Recommendations

• CPU: 2x AMD EPYC 7742 CPUs with hyper-threading
• Memory: >= 512GB
• Testing Storage: 8x Intel P4618 DC SSDs
• Data & System Storage: Any PCI-e attached SSD without MD RAIDs
• OS: Debian 11.1 with Linux kernel 5.10

System Setup

Clone TriCache

git clone -b osdi-ae https://github.com/thu-pacman/TriCache.git
cd TriCache
# add to ~/.bashrc
export TRICACHE_ROOT=$HOME/TriCache
git submodule update --init --recursive
cd ..

Install dependency

# Tested under Debian 11.1
# 	Boost: 1.74.0
# 	Thrift: 0.13.0
# 	TBB: 2020.3
sudo apt install vim-nox tmux rsync wget htop numactl time \
	xfsprogs psmisc
sudo apt install build-essential cmake git pkg-config libnuma-dev \
	libboost-all-dev libaio-dev libhwloc-dev libatlas-base-dev \
	zlib1g-dev thrift-compiler libthrift-dev libtbb-dev \
	libgflags-dev openjdk-11-jdk maven gdisk kexec-tools \
	python3 python3-matplotlib python3-numpy

# Install clang-13 from LLVM source
sudo apt install lsb-release software-properties-common
wget https://apt.llvm.org/llvm.sh
chmod +x llvm.sh
sudo ./llvm.sh 13
sudo apt install libomp-13-dev

# Build & install SPDK with TriCache's patch
git clone -b v21.04 https://github.com/spdk/spdk.git
cd spdk
git submodule update --init
git apply $TRICACHE_ROOT/deps/spdk.patch
sudo ./scripts/pkgdep.sh
./configure --prefix=$HOME/spdk-install --with-shared
make -j
make install -j
cd ..

# add to ~/.bashrc
export \
	PKG_CONFIG_PATH=$HOME/spdk-install/lib/pkgconfig:$PKG_CONFIG_PATH
export LD_LIBRARY_PATH=$HOME/spdk-install/lib:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=$HOME/spdk/dpdk/build/lib:$LD_LIBRARY_PATH
export SPDK_ROOT=$HOME/spdk

# testing SPDK examples (spdk_nvme_identify, spdk_nvme_perf)

# Install 4.14 kernel for FastMap
$TRICACHE_ROOT/install_4.14_kernel.sh

OS configurations

# Making sudo password-free
# add to /etc/sudoers 
%sudo   ALL=(ALL) NOPASSWD:ALL

# VM configures for SPDK and tunning swapping in-memory performance
# add following lines to /etc/sysctl.conf
vm.max_map_count=262144
vm.dirty_ratio=99
vm.dirty_background_ratio=99
vm.dirty_expire_centisecs=360000
vm.dirty_writeback_centisecs=360000
vm.swappiness=1

# ulimit configures for SPDK
# add following lines to /etc/security/limits.conf
* hard memlock unlimited
* soft memlock unlimited
* soft nofile 1048576

# disable transparent_hugepage for SPDK
# edit /etc/default/grub
# add transparent_hugepage=never to GRUB_CMDLINE_LINUX 
sudo update-grub

# Support key-based ssh to localhost
ssh-keygen
ssh-copy-id localhost
# test
ssh localhost

# reboot
sudo reboot

Build TriCache

cd $TRICACHE_ROOT
scripts/build.sh

Configure TriCache

Edit $TRICACHE_ROOT/scripts/config.sh to configure ssds for TriCache.

• SSDArray lists disk ids for testing disks. ls -lh /dev/disk/by-id will list all disk ids and their block id (nvmeXnY).
• SSDPCIe lists PCIe addresses testing disks. cat /sys/block/nvmeXnY/device/address will show PCIe the address.
• CACHE_16_SERVER_CONFIG lists disks and cores used by TriCache. Each item is formed like server-core-id,disk-pci-address,nsid,disk-offset.
• CACHE_16_SERVER_CORES lists cores used by TriCache.
• CACHE_32_SERVER_CONFIG and CACHE_32_SERVER_CORES a 32-server configurations.
• All the above disks will be formatted multiple times. YOU WILL LOSE THEIR DATA!
• All the above disks will be heavily written. THEY MAY BE BROKEN!
• Guide for selecting cores:
  1. Distribute the servers as evenly as possible among multiple numa-nodes
  2. Use hyperthreading to bind servers to as few physical cores as possible while avoiding over-subscribe for servers.
  3. For each server, find the closest SSDs binding to it.

Test TriCache

# setup SPDK, with 4GB DMA memory
$TRICACHE_ROOT/scripts/setup_spdk.sh 4

# "hello world" testing
$TRICACHE_ROOT/scripts/test_build.sh
# it should exit without error
# Reference log:
# Init 000 at 000 CPU, 00 Node
# Init 001 at 128 CPU, 00 Node
# Init 002 at 016 CPU, 01 Node
# Init 003 at 144 CPU, 01 Node
# Init 006 at 048 CPU, 03 Node
# Init 007 at 176 CPU, 03 Node
# Init 005 at 160 CPU, 02 Node
# Init 004 at 032 CPU, 02 Node
# Init 013 at 224 CPU, 06 Node
# Init 012 at 096 CPU, 06 Node
# Init 014 at 112 CPU, 07 Node
# Init 015 at 240 CPU, 07 Node
# Init 008 at 064 CPU, 04 Node
# Init 009 at 192 CPU, 04 Node
# Init 010 at 080 CPU, 05 Node
# Init 011 at 208 CPU, 05 Node
# Deconstructing SPDK Env

# reset devices from SPDK
$TRICACHE_ROOT/scripts/reset_spdk.sh

Generating and Preprocessing Datasets

• The experiments about graph processing use the uk-2014 dataset which is available here.
• The experiments about key-value stores use the mixgraph (prefix-dist) workload and the db_bench tool from RocksDB.
• The experiments about big-data analytics use terasort datasets generated by teragen from Hadoop.
• The experiments about graph database use LDBC SNB interactive benchmark generated by LDBC SNB Datagen.

The preprocessed datasets to reproduce the experimental results are available at: https://pacman.cs.tsinghua.edu.cn/public/TriCacheAEData.tar.zst, (MD5: 1560459277ffec508ff515dfbf931686).

# Unzip the datasets and put them into /mnt/data/TriCache
# /mnt/data/TriCache
# ├── flashgraph
# ├── ligra
# ├── livegraph
# ├── rocksdb
# └── terasort

sudo mkdir -p /mnt/data/TriCache
sudo tar --zstd -xf TriCache.tar.zst -C /mnt/data/TriCache

Reproduce Evaluations

• All the configured disked will be formatted multiple times. YOU WILL LOSE THEIR DATA!
• MAKE SURE there is no md raid like /dev/md*. MD RAIDS WILL BE BROKEN!
• /mnt/data/TriCache/temp, /mnt/raid and /mnt/ssd[0-15] will be used as temporary directories, please MAKE SURE no important data is stored in these directories.
• Please execute scripts in tmux or screen
• We recommend plotting the figures by following the later Plot Figures Section after running each part of the experiments.

Make an empty directory for storing logs, such as $HOME/results, and cd into it.

mkdir -p $HOME/results
cd $HOME/results

"Quick & Small" Experiments

It covers important cases in our evaluations. We think these results can support our claims within a short period.

# Run "Quick & Small" Experiments
# Taking about 7 hours
$TRICACHE_ROOT/scripts/run_all_small.sh

"One-click" Script for All Experiments

It combines all experiments (except FastMap and long-running parts) in one script. The following sections will describe how to run each experiment separately.

# Run "one-click" script
# Taking about 60 hours
$TRICACHE_ROOT/scripts/run_all.sh

Graph Processing (Section 4.1)

All-in-one

It combines all experiments (except long-running experiments) in one script.

# Taking about 11 hours 30 minutes
# It will process PageRank, WCC, BFS for uk-2014 dataset with 
# 	Ligra (TriCache), Ligra (Swapping), and FlashGraph.
# We move 64GB, 32GB, and 16GB with Ligra (Swapping)
# 	to Long-running Experiments
# Logs are placed in results_ligra_cache, results_ligra_swap, 
# 	and results_flashgraph
$TRICACHE_ROOT/scripts/run_all.sh graph_processing 

One-by-one

# Ligra (TriCache)
# Taking about 4 hours 30 minutes
# Logs are placed in results_ligra_cache
$TRICACHE_ROOT/scripts/run_all.sh graph_processing ligra_cache

# Ligra (Swapping)
# Taking about 5 hours 30 minutes
# Logs are placed in results_ligra_swap
$TRICACHE_ROOT/scripts/run_all.sh graph_processing ligra_swap

# FlashGraph
# Taking about 1 hour 30 minutes
# Logs are placed in results_flashgraph
$TRICACHE_ROOT/scripts/run_all.sh graph_processing flashgraph

Long-running Experiments

It will force to execute long-time experiments. These experiments will take more than 16 hours, so PLEASE skip them first.

# Ligra (Swapping) Long-running Parts
# Logs are placed in results_ligra_swap
$TRICACHE_ROOT/scripts/run_all.sh graph_processing ligra_swap_slow

Key-Value Stores (RocksDB: Section 4.2)

All-in-one

# Taking about 5 hours
# It will execute mixgraph workload with db_bench of RocksDB 
# 	PlainTable (TriCache), BlockBasedTable and PlainTable (mmap).
# We reduce some requests number to limit total execution time
# 	for long-running cases with BlockBasedTable and PlainTable (mmap)
# Logs are placed in results_rocksdb_cache, results_rocksdb_block, 
# 	and results_rocksdb_mmap
$TRICACHE_ROOT/scripts/run_all.sh rocksdb 

One-by-one

# PlainTable (TriCache)
# Taking about 2 hours
# Logs are placed in results_rocksdb_cache
$TRICACHE_ROOT/scripts/run_all.sh rocksdb rocksdb_cache

# BlockBasedTable
# Taking about 1 hour
# Logs are placed in results_rocksdb_block
$TRICACHE_ROOT/scripts/run_all.sh rocksdb rocksdb_block

# PlainTable (mmap)
# Taking about 2 hours
# Logs are placed in results_rocksdb_mmap
$TRICACHE_ROOT/scripts/run_all.sh rocksdb rocksdb_mmap

Big-Data Analytics (Terasort: Section 4.3)

All-in-one

# Taking about 14 hours
# It will execute 150GB and 400GB Terasort with 
# 	ShuffleSort (TriCache), GNUSort (TriCache), and Spark,
# 	and will execute 150GB dataset Terasort with
# 	ShuffleSort (Swapping), GNUSort (Swapping).
# We move the 400GB with ShuffleSort/GNUSort (Swapping)
# 	to Long-running Experiments
# Logs are placed in results_terasort_cache, results_terasort_swap, 
# 	and results_terasort_spark
$TRICACHE_ROOT/scripts/run_all.sh terasort 

One-by-one

# ShuffleSort (TriCache) and GNUSort (TriCache)
# Taking about 3 hours
# Logs are placed in results_terasort_cache
$TRICACHE_ROOT/scripts/run_all.sh terasort terasort_cache

# ShuffleSort (Swapping) and GNUSort (Swapping)
# Taking about 7 hours 30 minutes
# Logs are placed in results_terasort_swap
$TRICACHE_ROOT/scripts/run_all.sh terasort terasort_swap

# Spark
# Taking about 3 hours 30 minutes
# Logs are placed in results_terasort_spark
$TRICACHE_ROOT/scripts/run_all.sh terasort terasort_spark

Long-running Experiments

It will force to execute long-time experiments. These experiments will take more than 48 hours, so PLEASE skip them first.

# GNUSort (Swapping) Long-running Parts
# Logs are placed in results_terasort_swap
$TRICACHE_ROOT/scripts/run_all.sh terasort terasort_swap_slow

Graph Database (LiveGraph: Section 4.4)

All-in-one

# Taking about 8 hours
# It will execute SF30 and SF100 SNB Interactive benchmarks 
# 	with LiveGraph (TriCache) and LiveGraph (mmap).
# Logs are placed in results_livegraph_cache 
# 	and results_livegraph_mmap
$TRICACHE_ROOT/scripts/run_all.sh livegraph 

One-by-one

# LiveGraph (TriCache)
# Taking about 4 hours
# Logs are placed in results_livegraph_cache
$TRICACHE_ROOT/scripts/run_all.sh livegraph livegraph_cache

# LiveGraph (mmap)
# Taking about 4 hours
# Logs are placed in results_livegraph_mmap
$TRICACHE_ROOT/scripts/run_all.sh livegraph livegraph_mmap

Micro-benchmarks (Section 4.5)

All-in-one (Except FastMap)

# Taking about 11 hours
# Logs are placed in results_microbenchmark_cache 
# 	and results_microbenchmark_mmap
$TRICACHE_ROOT/scripts/run_all.sh microbenchmark 

One-by-one

# TriCache
# Taking about 6 hours
# Logs are placed in results_microbenchmark_cache 
$TRICACHE_ROOT/scripts/run_all.sh microbenchmark microbenchmark_cache

# mmap
# Taking about 5 hours
# Logs are placed in results_microbenchmark_mmap
$TRICACHE_ROOT/scripts/run_all.sh microbenchmark microbenchmark_mmap

FastMap Experiments

FastMap experiments run on an old kernel and require rebooting 20 times because it might crash the kernel. The scripts will automatically reboot the server. These experiments are very tricky and take about 9 hours. We recommend that you be prepared for unexpected cases, such as losing the server and requiring cold resets.

# Preparing disks for FastMap
$TRICACHE_ROOT/scripts/prepare_fastmap.sh

# Reboot to 4.14 kernel
$TRICACHE_ROOT/scripts/kexec.sh

# Wait rebooting, and ssh back again
# Check the kernel version is 4.14
uname -a

Go to another server to continue, such as the jumper.

# Download the driver script from 
wget https://raw.githubusercontent.com/thu-pacman/TriCache/osdi-ae\
/scripts/run_microbenchmark_fastmap_driver.sh

# Run the driver script
# Taking 9 hours
# Logs are placed in results_microbenchmark_fastmap
./run_microbenchmark_fastmap_driver.sh HOST RESULTS_DIR

# ssh back again 
# Check the kernel version is 5.10
uname -a

Performance Breakdown (Section 4.6)

All-in-one

# Taking about 9 hours
# We disable Shared-Cache-Only for GNU Terasort
# 	because it cannot finish within 5 hours, 
# 	and such low performance is already meaningless.
# Logs are placed in results_breakdown
$TRICACHE_ROOT/scripts/run_all.sh breakdown 

Plot Figures

Make an empty directory for storing figures, such as $HOME/figures, and cd into it.

mkdir -p $HOME/figures
cd $HOME/figures

"Quick & Small" Experiments

# After running "quick & small" experiments
$TRICACHE_ROOT/scripts/plot_all.sh small RESULTS_PATH

All Experiments

# After running experiments (with/without long-running experiments)
$TRICACHE_ROOT/scripts/plot_all.sh RESULTS_PATH

Graph Processing (Section 4.1 and Figure 8)

# After running graph processing experiments
#	(with/without long-running experiments)
$TRICACHE_ROOT/scripts/plot_all.sh graph_processing RESULTS_PATH

Key-Value Stores (Section 4.2 and Figure 9)

# After running key-value experiments
$TRICACHE_ROOT/scripts/plot_all.sh rocksdb RESULTS_PATH

Big-Data Analytics (Section 4.3 and Figure 10)

# After running big-data analytics experiments
#	(with/without long-running experiments)
$TRICACHE_ROOT/scripts/plot_all.sh terasort RESULTS_PATH

Graph Database (Section 4.4 and Figure 11)

# After running graph database experiments
$TRICACHE_ROOT/scripts/plot_all.sh livegraph RESULTS_PATH

Micro-benchmarks (Section 4.5 and Figure 12)

# After running micro-benchmarks experiments
#	(with/without FastMap experiments)
$TRICACHE_ROOT/scripts/plot_all.sh microbenchmark RESULTS_PATH

Performance Breakdown (Section 4.6 and Table 1)

# After running performance breakdown experiments
$TRICACHE_ROOT/scripts/plot_all.sh breakdown RESULTS_PATH