Hope that you'd be glad to add a star if you think this repo is helpful!

NTSocks develops an ultra-fast and compatible PCIe interconnect targetting for rack-scale disaggregations, leveraging routable PCIe fabric (e.g., Non Transparent Bridge -- NTB). Compared to state-of-the-art RDMA and user-space network stack (e.g., mTCP), such routable PCIe fabric achieves lower nanoseceond-level latency, and higher throughput close to PCIe bandwidth limits, due to eliminating the protocol translation overhead (between PCIe and network protocols) and complex in-NIC resource management (e.g., limited NIC cache for RDMA connection context and memory mapping table). This potentially meets the increasing demand of high-speed networking required by rack-scale disaggregation. NTSocks is thereby motivated as a software indirection layer to efficiently virtualize the PCIe transport resources for rack-scale disaggregated applications. At a high level, the NTSocks layer gets socket API calls from apps and process them by interacting with a user-level runtime library on behalf of the app, as the following architecture figure shows. NTSocks consists of three components: NTSocks library, Monitor, and Proxy. The runtime NTSocks library, called libnts, is located inside the application process and provides user-friendly and performant socket abstractions for applications. It is responsible for parsing and fowarding control and data path operations into NTSocks monitor and proxy, respectively. The control plane component monitor, called NTM, performs control path operations (such as connection setup when connect()). The data plane component proxy, called NTP, virtualizes global PCIe fabric transport resource (such as NTB mapped memory) in a lightweight manner, and then coordinates with libnts to perform fast data path operations (e.g., write()/read()). These three components communicate via shared memory. This repo contains implementation of NTSocks system. Please refer to each README under those subdirectories for more informations.

Note that our NTSocks work has been published at ACM CoNEXT 2022 -- Best Paper Award.

The libnts directory contains the implementation of NTSocks runtime library, including control (e.g., nts_connect()) and data path interfaces (e.g., nts_send()). This libnts library is located inside application process, which is responsible for parsing and forwarding control and data path requests into NTSocks monitor and proxy, respectively.

The monitor directory contains the implementation of NTSocks control plane component, including initilizing PCIe fabric routing information, allocating the allocation between connection and point-to-point PCIe interconnect, connection establishment over routable PCIe interconnect, and destroying PCIe fabric resources when disconnecting with remote sides.

The ntb-proxy directory contains the implementation of NTSocks data plane component, including partitioning PCIe transport resources with multi-cores for scaling up, batching tx/rx messages foarding, and hierarchical performance isolation mechanism.

The ./apps directory contains an NTSocks-enhanced key-value store system with varying YCSB workloads. Wherea the ./benchmark directory contains a set of function tests to evaluate NTSocks control and data plane interfaces.

Please refer to ./build-nts.sh script to build three NTSocks components. This will produce:

• libnts: libnts.so and libnts.a in libnts/lib/ directory; libnts.h, nts_config.h, and socket.h in libnts/include/ directory.
• NTM: An executable ntb-monitor binary in monitor/bin/ directory.
• NTP: An executable ntb-proxy binary in ntb-proxy/src/ directory.

Before executation, please ensure the PCIe NTB is enabled in BIOS and physical machines, while correctly setuping DPDK NTB Poll-Mode Driver.

1. Startup NTSocks control plane monitor (refer to ./monitor.sh).
2. Startup NTSocks data plane proxy (refer to ./proxy.sh).
3. Assuming to use LD_PRELOAD runtime library libnts, a socket based server program (refer to benchmark/unified_benchmark/ntb-server) and a client program (refer to benchmark/unified_benchmark/ntb-client) are executed without any code changes. They can communciate with each other transparently.

The code is released under the Apache 2.0 License.

This work was done collobratively with Intel DPDK team. Thanks to Intel for the PCIe NTB testbed and technical support!

If you use our NTSocks or related codes in your research, please cite our paper:

@inproceedings{huang2022ultra,
  title={An ultra-low latency and compatible PCIe interconnect for rack-scale communication},
  author={Huang, Yibo and Huang, Yukai and Yan, Ming and Hu, Jiayu and Liang, Cunming and Xu, Yang and Zou, Wenxiong and Zhang, Yiming and Zhang, Rui and Huang, Chunpu and others},
  booktitle={Proceedings of the 18th International Conference on emerging Networking EXperiments and Technologies},
  pages={232--244},
  year={2022}
}