• CMake
• Anaconda3
• Python packages: pip install python-memcached fabric
• Other dependencies: sudo apt install libmemcached-dev memcached libboost-all-dev libhiredis-dev
• Redis++:

  git clone https://github.com/sewenew/redis-plus-plus.git
  cd redis-plus-plus
  mkdir build
  cd build
  cmake ..
  make -j 8
  sudo make install

• Modify /etc/memcached.conf

  -l 0.0.0.0
  -I 128m
  -m 2048
  

We provide a script for downloading Twitter workloads in workloads/download_workloads.sh. Execute cd workloads && ./download_workloads.sh twitter to download and setup the execution directories for the Twitter workloads. You may at leat 1TB disk space to download and decompress these workloads.

We support executing client threads on multiple nodes. Different nodes are synchronized through memcached with a controller. We currently support two modes of executions:

• Separate data loading and workload execution
  • Load data:
    1. python3 scripts/ycsb_load_or_trans.py <memcached-ip> <num-clients>. The <num-clients> field should be the total number of client threads.
    2. On another client node: cd build && ./redis_load_ycsb <node-id> <num-local-threads> <num-total-threads> <memcached-ip> load <redis-ip> <execution-time> <output-fname> <mode>. The load and <execution-time> are a placeholders, taking no effect to the command. The <mode> field can either be cluster, single, or dmc_cluster. Note that this program will automatically read the workload file in ../workloads/ycsb, hence must be executed in the build directory. An output will be written to build/results as a json file.
  • Executing workload normally:
    1. python3 scripts/ycsb_load_or_trans.py <memcached-ip> <num-clients>
    2. On another client node: cd build && ./redis_trans_ycsb <node-id> <num-local-threads> <num-total-threads> <memcached-ip> <workload> <redis-ip> <execution-time> <output-fname> <mode> normal. The <workload> should be in [ycsba, ycsbb, ycsbc, ycsbd, twitter-compute, twitter-storage, twitter-transient] and The <execution-time> controls how long each thread iterates with the workload. The <mode> field can either be cluster, single, or dmc_cluster. Note that this program will automatically read the workload file in ../workloads/ycsb, hence must be executed in the build directory. An output will be written to build/results as a json file.
  • Executing workload for dmc cluster: 0. Prepare experiment settings: - In scripts/ycsb_dmc_cluster_elasticity.py there are two variables to control the execution time of scale out and scale in nodes, which are seconds_before_scale and seconds_before_shrink. - In dmc_cluster_config.json set num_initial_servers and num_scale_servers to control the number of servers used before and after resource scaling.
    1. python scripts/ycsb_dmc_cluster_elasticity.py <memcached-ip> <num-clients>.
    2. On another client node: cd build && ./redis_trans_ycsb <node-id> <num-local-threads> <num-total-threads> <memcached-ip> <workload> <redis-ip> <execution-time> <output-fname> <mode> elasticity. The <workload> should be in [ycsba, ycsbb, ycsbc, ycsbd, twitter-compute, twitter-storage, twitter-transient] and The <execution-time> controls how long each thread iterates with the workload. The <mode> field can either be cluster, single, or dmc_cluster. Note that this program will automatically read the workload file in ../workloads/ycsb, hence must be executed in the build directory. An output will be written to build/results as a json file.
• End-to-end testing (load data and execute the workload):
  1. python3 scripts/ycsb_load_and_trans.py <memcached-ip> <num-clients>. The <num-clients> field should be the total number of client threads.
  2. On another client node: cd build && ./redis_perf <node-id> <num-local-threads> <num-total-threads> <memcached-ip> <workload> <redis-ip> <execution-time> <output-fname> <mode> <experiment>. The <workload> should be in [ycsba, ycsbb, ycsbc, ycsbd] and The <execution-time> controls how long each thread iterates with the workload. The <mode> field can either be cluster, single, dmc_cluster, sentinel, or dmc_sentinel. Note that this program will automatically read the workload file in ../workloads/ycsb, hence must be executed in the build directory. An output will be written to build/results as a json file. The <experiment> field can either be normal, elasticity, or recovery.
• Redis elasticity test:
  • Horizontal Scaling 0. In scripts/redis-cluster-scale-horizontal.py, modify the two variables to control the execution time seconds_before_scale and seconds_before_shrink.
    1. python3 scripts/redis-cluster-scale-horizontal.py <num_initial_servers> <num_redis_clients> <memcached_ip>
    2. On another client node: cd build && ./redis_perf <node-id> <num-local-threads> <num-total-threads> <memcached-ip> <workload> <redis-ip> <execution-time> <output-fname> cluster elasticity. The <workload> should be in [ycsba, ycsbb, ycsbc, ycsbd] and The <execution-time> controls how long each thread iterates with the workload. Note that this program will automatically read the workload file in ../workloads/ycsb, hence must be executed in the build directory. An output will be written to build/results as a json file.
  • Vertical Scaling 0. In scripts/redis-cluster-scale-vertical.py, modify the two variables to control the execution time seconds_before_scale and seconds_before_shrink.
    1. python3 scripts/redis-cluster-scale-horizontal.py <num_redis_clients> <memcached_ip>
    2. On another client node: cd build && ./redis_perf <node-id> <num-local-threads> <num-total-threads> <memcached-ip> <workload> <redis-ip> <execution-time> <output-fname> cluster elasticity. The <workload> should be in [ycsba, ycsbb, ycsbc, ycsbd] and The <execution-time> controls how long each thread iterates with the workload. Note that this program will automatically read the workload file in ../workloads/ycsb, hence must be executed in the build directory. An output will be written to build/results as a json file.
• DMC Sentinel test:
  • Compute node failures 0. Prepare experiment settings: - Start 2 DMC servers one primary and one backup. - In scripts/ycsb_dmc_sentinel.py there are two variables to control the execution time of fail failing the primary node (seconds_before_fail) and the time for recover the backup node (seconds_before_rereplicate). - In dmc_cluster_config.json set num_initial_servers = 2 and num_scale_servers = 3. The IPs of DMC servers should also be set correctly. - Start a new backup during the experiment after the seconds_before_fail.
    1. python scripts/ycsb_dmc_sentinel.py <memcached-ip> <num-clients>.
    2. On another client node: cd build && ./redis_perf <node-id> <num-local-threads> <num-total-threads> <memcached-ip> <workload> <redis-ip> <execution-time> <output-fname> dmc_sentinel recovery. The <workload> should be in [ycsba, ycsbb, ycsbc, ycsbd, twitter-compute, twitter-storage, twitter-transient] and The <execution-time> controls how long each thread iterates with the workload. Note that this program will automatically read the workload file in ../workloads/ycsb, hence must be executed in the build directory. An output will be written to build/results as a json file.
  • Memory node failures 0. Prepare experiment settings: - Start 2 DMC servers one primary and one backup. - In dmc_cluster_config.json set num_initial_servers = 2 and num_scale_servers = 2. The IPs of DMC servers should also be set correctly. - Kill a memory node and rereplicate its memory regions.
    1. python scripts/ycsb_dmc_sentinel.py <memcached-ip> <num-clients>.
    2. On another client node: cd build && ./redis_perf <node-id> <num-local-threads> <num-total-threads> <memcached-ip> <workload> <redis-ip> <execution-time> <output-fname> dmc_sentinel recovery. The <workload> should be in [ycsba, ycsbb, ycsbc, ycsbd, twitter-compute, twitter-storage, twitter-transient] and The <execution-time> controls how long each thread iterates with the workload. Note that this program will automatically read the workload file in ../workloads/ycsb, hence must be executed in the build directory. An output will be written to build/results as a json file.

Note 1: In all above commands, the <redis-ip> field should be a URI with a scheme, e.g., tcp://127.0.0.1:7000.
Note 2: Create a results directory under the build directory. Otherwise, results can be lost.
Note 3: When executing Twitter workloads, we do not load data first.
Note 4: For dmc_cluster mode, we use dmc_cluster_config.json to setup nodes in the cluster and the number of initial servers and the number of scale servers (for elasticity experiment).
Note 5: For Redis elasticity test, we use the <execution-time> of redis-perf to control the time for executing the workload.

{
    "cont_tpt": [num_ops, num_ops, num_ops, ...], // record every 500ms
    "cont_lat_map": [[[lat_us, time], [lat_us, time], ...], 
                     [[lat_us, time], [lat_us, time], ...],
                     ...], //record every 500ms
}

• Working on redis-instance-starter.py to start a single redis server instance or a redis cluster with multiple nodes
  • Single server mode
  • Cluster mode with more than 3 nodes
  • Cluster mode with less than 3 nodes