CHIME: A Cache-Efficient and High-Performance Hybrid Index on Disaggregated Memory

This is the implementation repository of our SOSP'24 paper: CHIME: A Cache-Efficient and High-Performance Hybrid Index on Disaggregated Memory. This artifact provides the source code of CHIME and scripts to reproduce all the experiment results in our paper. CHIME, a Cache-efficient and High-performance Hybrid Index on disaggregated MEmory, combines the B+ tree with hopscotch hashing, to break the trade-off between cache consumption and read amplifications for existing DM range indexes.

• CHIME
  • Supported Platform
  • Create Cluster
  • Source Code (Artifacts Available)
  • Environment Setup
  • YCSB Workloads
  • Getting Started (Artifacts Functional)
  • Reproduce All Experiment Results (Results Reproduced)

Supported Platform

We strongly recommend you to run CHIME using the r650 instances on CloudLab as the code has been thoroughly tested there. We haven't done any test in other hardware environment.

If you want to reproduce the results in the paper, 10 r650 machines are needed; otherwise, fewer machines (i.e., 3) is OK. Each r650 machine has two 36-core Intel Xeon CPUs, 256 GB of DRAM, and one 100 Gbps Mellanox ConnectX-6 RNIC. Each RNIC is connected to a 100 Gbps Ethernet switch.

Create Cluster

Follow the steps below to create an experimental cluster with 10 r650 nodes on CloudLab:

1. Log into your own account.

2. Click Experiments|-->Create Experiment Profile. Upload ./script/cloudlab.profile provided in this repo. Input a file name (e.g., CHIME) and click Create to generate the experiment profile for CHIME.

3. Click Instantiate to create a 10-node cluster using the profile (This takes about 7 minutes).

4. Try logging into and check each node using the SSH commands provided in the List View on CloudLab. If you find some nodes have broken shells, you can reload them via List View|-->Reload Selected.

Source Code (Artifacts Available)

Now you can log into all the CloudLab nodes. Using the following command to clone this github repo in the home directory (e.g., /users/TempUser) of all nodes:

git clone https://github.com/dmemsys/CHIME.git

Environment Setup

You have to install the necessary dependencies in order to build CHIME. Note that you should run the following steps on all nodes you have created.

1. Set bash as the default shell. And enter the CHIME directory.

   sudo su
   chsh -s /bin/bash
   cd CHIME

2. Install Mellanox OFED.

   # It doesn't matter to see "Failed to update Firmware"
   # This takes about 8 minutes
   sh ./script/installMLNX.sh

3. Resize disk partition.

   Since the r650 nodes remain a large unallocated disk partition by default, you should resize the disk partition using the following command:

   # It doesn't matter to see "Failed to remove partition" or "Failed to update system information"
   sh ./script/resizePartition.sh
   # This takes about 6 minutes
   reboot
   # After rebooting, log into all nodes again and execute:
   sudo su
   resize2fs /dev/sda1

4. Enter the directory. Install the required libraries.

   cd CHIME
   # This takes about 3 minutes
   sh ./script/installLibs.sh

YCSB Workloads

You should run the following steps on all nodes.

1. Download YCSB source code.

   sudo su
   cd CHIME/ycsb
   curl -O --location https://github.com/brianfrankcooper/YCSB/releases/download/0.11.0/ycsb-0.11.0.tar.gz
   tar xfvz ycsb-0.11.0.tar.gz
   mv ycsb-0.11.0 YCSB

2. We first generate a small set of YCSB workloads here for a quick start (i.e, kick-the-tires).

   # This takes about 20 seconds
   sh generate_small_workloads.sh

Getting Started (Artifacts Functional)

• Hugepages setting on all nodes (This step should be re-execute every time you reboot the nodes).

  sudo su
  echo 36864 > /proc/sys/vm/nr_hugepages
  ulimit -l unlimited

• Return to the root directory of CHIME and execute the following commands on all nodes to compile CHIME:

  mkdir build; cd build; cmake ..; make -j

• Execute the following command on one node to initialize the memcached:

  /bin/bash ../script/restartMemc.sh

• Execute the following command on all nodes to split the workloads:

  python3 ../ycsb/split_workload.py <workload_name> <key_type> <CN_num> <client_num_per_CN>

  • workload_name: the name of the workload to test (e.g., a / b / c / d / e).
  • key_type: the type of key to test (i.e., randint).
  • CN_num: the number of CNs.
  • client_num_per_CN: the number of clients on each CN.

  Example:

  python3 ../ycsb/split_workload.py a randint 10 24

• Execute the following command on all nodes to conduct a YCSB evaluation:

  ./ycsb_test <CN_num> <client_num_per_CN> <coro_num_per_client> <key_type> <workload_name>

  • coro_num_per_client: the number of coroutine in each client (2 is recommended).

  Example:

  ./ycsb_test 10 24 2 randint a

• Results:

  • Throughput: the throughput of CHIME of the whole cluster will be shown in the terminal of the first node (with 10 epoches by default).

  • Latency: execute the following command on one node to calculate the latency results of the whole cluster:

    python3 ../us_lat/cluster_latency.py <CN_num> <epoch_start> <epoch_num>

    Example:

    python3 ../us_lat/cluster_latency.py 10 1 10

Reproduce All Experiment Results (Results Reproduced)

We provide code and scripts in ./exp folder for reproducing our experiments. For more details, see ./exp/README.md.