Welcome to Postman Project
Postman is a distributed service aming to improve system performance in the network layer for small packets by offloading packet processing.
Introduction
Network layer usually performs poorly when processing small packets, because there is a constant overhead associated with processing every packet. For example, for each packet, Linux needs to handle the interrupt (and the associated context switch) caused by the packet, perform protocol operations, such as checksum verification and congestion control, and process related system calls (also the associated context switch) made by applications. As a result, when transferring 64-byte packets on 10Gb ethernet, Linux can only deliver a throughput of 0.5Gb/s.
Postman deployes a number of helper nodes, which are equipped with new techniques for packet processing (e.g., DPDK), to assemble small packets to the same server, so that a server application can enjoy both the high performance of processing large packets and the convenience of working on a classic operating system.
Postman Design
Postman aims to achieve the following targets:
-
Efficient packet processing.
A helper node needs to assemble small packets efficiently. For efficiency, Postman implements a helper node under the popular Data Plane Development Kit (DPDK) framework. To further improve performance, Postman 1) removes redundant information in packets’ headers when assembling them, so that a server application does not need to pay the bandwidth cost for headers; 2) strikes a balance between latency and throughput by dynamically comparing the additional delay introduced by packets assembly and the reduced queueing delay. -
Transparent offloading.
To make Postman practical, packet redirection, assembly, and disassembly should be transparent to applications. To achieve that, we build a virtual connection library (VCL), providing an illusion to a client that it is still talking to the server directly. VCL provides the same interface as classic TCP libraries, but internally, a client will re-direct packets to helper nodes, and a server will disassemble large packets from helper nodes. Furthermore, to tolerate helper node failures, VCL incorporates a connection re-establishment protocol, guaranteeing that packets will not be lost, reordered, or duplicated, despite helper node failures. -
Latency-aware helper placement and load balancing.
Postman manages a helper as an instance of Virtual Network Function (VNF) and use the recent Software Defined Networking (SDN) and Network Functions Virtualization (NFV) techniques to design an automatic framework to place helpers in different locations of the network, such that network management is simplified and network latency is minimized