Freeflow is a high performance container overlay network that enables RDMA communication and accelerates TCP socket to the same as the host network.
Freeflow works on top of popular overlay network solutions including Flannel, Weave, etc. The containers have their individual virtual network interfaces and IP addresses, and do not need direct access to the hardware NIC interface. A lightweight Freeflow library inside containers intercepts RDMA and TCP socket APIs, and a Freeflow router outside containers helps accelerate those APIs.
Freeflow is developed based on Linux RDMA project (https://github.com/linux-rdma/rdma-core), and released with MIT license.
Freeflow works in three modes: fully-isolated RDMA (master branch), semi-isolated RDMA, and TCP (tcp branch).
Fully-isolated RDMA provides the best isolation between different containers and works the best in multi-tenant environment, e.g., cloud. While it offers typical RDMA performance (40Gbps throughput and 1 or 2 microsecond latency), this comes with some CPU overhead penalty.
The TCP mode accelerates the TCP socket performance to the same as host network. On a typical Linux server with a 40Gbps NIC, it can achieve 25Gbps throughput for a single TCP connection and less than 20 microsecond latency.
We will release semi-isolated RDMA in the future. It has the same CPU efficiency as host RDMA, while does not have full isolation on the data path. It works the best for single-tenant clusters, e.g., an internal cluster.
Below show the performance of Spark and Tensorflow running in fully-isolated RDMA mode on servers connected with 40Gbps RDMA network.
Below are the steps of running Freeflow in fully-isolated RDMA mode. For TCP mode, refer to the README in tcp branch.
Step 1: Start Freeflow router (one instance per server)
sudo docker run --name router1 --net host -e "FFR_NAME=router1" -e "LD_LIBRARY_PATH=/usr/lib/:/usr/local/lib/:/usr/lib64/" -v /sys/class/:/sys/class/ -v /freeflow:/freeflow -v /dev/:/dev/ --privileged -it -d ubuntu:14.04 /bin/bash
Then log into the router container with
sudo docker exec -it router1 bash
Download and install RDMA libraries and drivers in router container. Currently, Freeflow is developed and tested with "MLNX_OFED_LINUX-4.0-2.0.0.1-ubuntu14.04-x86_64.tgz" You can download it from http://www.mellanox.com/page/products_dyn?product_family=26.
Then, build the code using the script build-router.sh. In ffrouter/, start the router by running "./router router1".
Step 2: Repeat Step 1 to start the router in other hosts. You can capture a Docker image of router1 for avoiding repeating the installations and building.
Step 3: Start an application container on the same host as router1
sudo docker run --name node1 --net weave -e "FFR_NAME=router1" -e "FFR_ID=10" -e "LD_LIBRARY_PATH=/usr/lib" -e --ipc=container:router1 -v /sys/class/:/sys/class/ -v /freeflow:/freeflow -v /dev/:/dev/ --privileged --device=/dev/infiniband/uverbs0 --device=/dev/infiniband/rdma_cm -it -d ubuntu /bin/bash
You may use any container overlay solution. In this example, we use Weave (https://github.com/weaveworks/weave).
Environment variable "FFR_NAME=router1" points to the container to the router (router1) on the same host; "FFR_ID=10" is the ID of the contaienr in FreeFlow. Each container on the same host should have a unique FFR_ID. We are removing FFR_ID in next version.
Download and install the same version of RDMA libraries and drivers as Step 1. Then build the the code of libraries/ and libmempool/ and install to /usr/lib/ (which is default).
Step 4: Repeat Step 3 to start customer containers in more hosts. You can capture a Docker image of node1 for avoiding repeating the installations and building.
Attention: the released implementation hard-codes the host IPs and virtual IP to host IP mapping in https://github.com/Microsoft/Freeflow/blob/master/ffrouter/ffrouter.cpp#L215 and https://github.com/Microsoft/Freeflow/blob/master/ffrouter/ffrouter.h#L76. For quick tests, you can edit it according to your environment. Ideally, the router should read it from container overlay controller/zookeeper/etcd.
Validation: in customer containers, install RDMA perftest tools with "sudo apt-get install perftest". Try "ib_send_bw" or "ib_send_lat".
For RDMA, Freeflow has been tested with RDMA-based Spark (http://hibd.cse.ohio-state.edu/), HERD (https://github.com/efficient/HERD), Tensorflow with RDMA enabled (https://github.com/tensorflow/tensorflow) and rsocket (https://linux.die.net/man/7/rsocket). Most RDMA applications should run with no (or very little) modification, and outperform traditional TCP socket-based implementation.
For TCP, Freeflow has also been tested with many applications/framework, including DLWorkspace (https://github.com/Microsoft/DLWorkspace), Horovod (https://github.com/uber/horovod), Memcached, Nginx, PostSQL, and Kafka.
This implementation is a research prototype that shows the feasibility. It is NOT production quality code. The technical details will be published in academic papers. If you have any questions, please raise issues on Github or contact the authors below.
Hongqiang Harry Liu ([email protected])