/eRPC

Efficient RPCs for datacenter networks

Primary LanguageC++OtherNOASSERTION

News [Dec 2023]: Check out Machnet for fast and easy DPDK-based messaging. It supports many types of cloud VMs and bare-metal NICs, multiple application processes, language bindings, etc.

eRPC is a fast and general-purpose RPC library for datacenter networks. Our NSDI 2019 paper describes the system in detail. Documentation can be generated by running doxygen.

Some highlights:

  • Multiple supported networks: Ethernet, InfiniBand, and RoCE
  • Low latency: 2.3 microseconds round-trip RPC latency with UDP over Ethernet
  • Performance for small 32-byte RPCs: ~10M RPCs/sec with one CPU core, 60--80M RPCs/sec with one NIC.
  • Bandwidth for large RPC: 75 Gbps on one connection (one CPU core at server and client) for 8 MB RPCs
  • Scalability: 20000 RPC sessions per server
  • End-to-end congestion control that tolerates 100-way incasts
  • Nested RPCs, and long-running background RPCs
  • A port of Raft as an example. Our 3-way replication latency is 5.3 microseconds with traditional UDP over Ethernet.

System requirements

  • NICs: Fast (10 GbE+) NICs are needed for good performance. eRPC works best with Mellanox Ethernet and InfiniBand NICs. Any DPDK-capable NICs also work well.
  • System configuration:
    • At least 1024 huge pages on every NUMA node, and unlimited SHM limits
    • On a machine with n eRPC processes, eRPC uses kernel UDP ports {31850, ..., 31850 + n - 1}. These ports should be open on the management network. See scripts/firewalld/erpc_firewall.sh for systems running firewalld.

eRPC quickstart

  • Build and run the test suite: cmake . -DPERF=OFF -DTRANSPORT=dpdk; make -j; sudo ctest.
    • DPERF=OFF enables debugging, which greatly reduces performance. Set DPERF=ON for good performance.
    • Here, dpdk should be replaced with infiniband for InfiniBand NICs.
    • A machine with two ports is needed to run the unit tests if DPDK is chosen. Run scripts/run-tests-dpdk.sh instead of ctest.
  • Run the hello_world application:
    • cd hello_world
    • Edit the server and client hostnames in common.h
    • Based on the transport that eRPC was compiled for, compile hello_world using make dpdk, or make infiniband.
    • Run ./server at the server, and ./client at the client
  • Generate the documentation: doxygen

Supported bare-metal NICs:

  • Ethernet/UDP mode:
    • DPDK-enabled bare-metal NICs: Use DTRANSPORT=dpdk. We have primarily tested Mellanox CX3--CX5 NICs.
    • DPDK-enabled NICs on Microsoft Azure: Use -DTRANSPORT=dpdk -DAZURE=on
  • RDMA (InfiniBand/RoCE) NICs: Use DTRANSPORT=infiniband. Add DROCE=on if using RoCE.

Running eRPC over DPDK on Microsoft Azure VMs

  • eRPC works well on Azure VMs with accelerated networking.

  • Configure two Ubuntu 18.04 VMs as below. Use the same resource group and availability zone for both VMs.

    • Uncheck "Accelerated Networking" when launching each VM from the Azure portal (e.g., F32s-v2). For now, this VM should have just the control network (i.e., eth0) and lo interfaces.
    • Add a NIC to Azure via the Azure CLI: az network nic create --resource-group <your resource group> --name <a name for the NIC> --vnet-name <name of the VMs' virtual network> --subnet default --accelerated-networking true --subscription <Azure subscription, if any> --location <the VM's availability zone>
    • Stop the VM launched earlier, and attach the NIC created in the previous step to the VM (i.e., in "Networking" -> "Attach network interface").
    • Re-start the VM. It should have a new interface called eth1, which eRPC will use for DPDK traffic.
  • Prepare DPDK 21.11

    • rdma-core must be installed from source. We recommend the tag stable-v40. First, install its dependencies listed in rdma-core's README. Then, in the rdma-core` directory:
      • cmake .
      • sudo make install
    • Install upstream pre-requisite libraries and modules:
      • sudo apt install make cmake g++ gcc libnuma-dev libgflags-dev numactl
      • sudo modprobe ib_uverbs
      • sudo modprobe mlx4_ib
    • Download the DPDK tarball and extract it. Other DPDK versions are not supported.
    • Edit config/common_base by changing CONFIG_RTE_LIBRTE_MLX5_PMD and CONFIG_RTE_LIBRTE_MLX4_PMD to y instead of n.
    • Build and locally install DPDK:
export RTE_SDK=<some dpdk directory>
git clone --depth 1 --branch 'v21.11' https://github.com/DPDK/dpdk.git "${RTE_SDK}"
cd "${RTE_SDK}"
meson build -Dexamples='' -Denable_kmods=false -Dtests=false -Ddisable_drivers='raw/*,crypto/*,baseband/*,dma/*'
cd build/
DESTDIR="${RTE_SDK}/build/install" ninja install
  • Create hugepages:
sudo bash -c "echo 2048 > /sys/devices/system/node/node0/hugepages/hugepages-2048kB/nr_hugepages"
sudo mkdir /mnt/huge
sudo mount -t hugetlbfs nodev /mnt/huge
  • Build eRPC's library and latency benchmark:
cmake . -DTRANSPORT=dpdk -DAZURE=on
make
make latency
  • Create the file scripts/autorun_process_file like below. Here, do not use the IP addresses of the accelerated NIC (i.e., not of eth1).
<Public IPv4 address of VM #1> 31850 0
<Public IPv4 address of VM #2> 31850 0
  • Run the eRPC application (the latency benchmark by default):
    • At VM #1: ./scripts/do.sh 0 0
    • At VM #2: ./scripts/do.sh 1 0

Configuring and running the provided benchmarks

  • The apps directory contains a suite of benchmarks and examples. The instructions below are for this suite of applications. eRPC can also be simply linked as a library instead (see hello_world/ for an example).
  • To build an application, create scripts/autorun_app_file and change its contents to one of the available directory names in apps/. See scripts/example_autorun_app_file for an example. Then generate a Makefile using cmake . -DTRANSPORT=dpdk/infiniband.
  • Each application directory in apps/ contains a config file that must specify all flags defined in apps/apps_common.h. For example, num_processes specifies the total number of eRPC processes in the cluster.
  • The URIs of eRPC processes in the cluster are specified in scripts/autorun_process_file. Each line in this file must be <hostname> <management udp port> <numa_node>.
  • Run scripts/do.sh for each process:
    • With single-CPU machines: num_processes machines are needed. Run scripts/do.sh <i> 0 on machine i in {0, ..., num_processes - 1}.
    • With dual-CPU machines: num_machines = ceil(num_processes / 2) machines are needed. Run scripts/do.sh <i> <i % 2> on machine i in {0, ..., num_machines - 1}.
  • To automatically run an application at all processes in scripts/autorun_process_file, run scripts/run-all.sh. For some applications, statistics generated in a run can be collected and processed using scripts/proc-out.sh.

Getting help

  • GitHub issues are preferred over email. Please include the following information in the issue:
    • NIC model
    • rdma_core version and DPDK version
    • Operating system

Contact

Anuj Kalia

License

	Copyright 2018, Carnegie Mellon University

    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at

        http://www.apache.org/licenses/LICENSE-2.0

    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    See the License for the specific language governing permissions and
    limitations under the License.