In-Network Aggregation (INA) offloads the gradient aggregation in distributed training (DT) onto programmable switches, where the switch memory could be allocated to jobs in either synchronous or statistical multiplexing mode. Statistical INA has advantages in switch memory utilization, control-plane simplicity, and management safety, but it faces the problem of cross-layer resource efficiency in job placement. This paper presents a job placement system NetPack for clusters with statistical INA, which aims to maximize the utilization of both computation and network resources. NetPack periodically batches and places jobs into the cluster. When placing a job, NetPack runs a steady state estimation algorithm to acquire the available resources in the cluster, heuristically values each server according to its available resources (GPU and bandwidth), and runs a dynamic programming algorithm to efficiently search for servers with the highest value for the job. Our prototype of NetPack and the experiments demonstrate that NetPack outperforms prior job placement methods by 45% in terms of average job completion time on production traces.

Accurate link latency information is essential for various traffic engineering problems, such as routing design and network diagnosis. However, due to the routing constraints, link latency measurement is still a challenging problem in hybrid software-defined networks (SDNs) and little literature has been found in this field. Due to cost, a hybrid SDN composed of conventional routers and SDN switches will exist for some time to come, which shows the necessity and the urgency to address this problem. In this paper, we investigate this link latency measurement problem in two different scenarios: 1) the conventional routers can only support the shortest path routing protocol, and 2) the conventional routers can support source routing protocol. For both of these scenarios, we show how to deploy a minimum number of monitors and how to construct measurement paths between monitors to measure all the link latencies. Several algorithms are presented to solve these problems and the evaluations on different topologies prove the superiority of the proposed methods.

Accurate link latency information is required for solving many traffic engineering problems, such as congestion avoidance routing and load balancing. Many link latency measurement schemes have been proposed for Software-Defined Networks (SDNs), and many of these require the SDN controllers to inject and remove entries into and from flow tables dynamically. If done frequently, this process can generate a large amount of traffic overhead and consume a lot of processing and computation resources. In this paper, we present a link latency measurement scheme, where entries for latency measurement are already embedded inside flow tables. The main challenge in designing such a scheme lies in how to select measurement paths that can minimize the number of entries added to the expensive TCAM-based flow tables and the amount of traffic generated for latency measurement. We present solutions that can minimize both in this paper.

Accurate traffic matrix estimation is critical to solving many networking problems, such as routing and network provisioning, etc. This task in a traditional IP network is tackled with link load measurements. But the accuracy of such an approach is low because the underlying system of linear equations governing the traffic estimation problem is highly under-determined in this approach. A software-defined network (SDN) provides measurements of more types of flows, and thus opens up new opportunities for tackling the traffic matrix estimation problem. Furthermore, the controller of an SDN can dynamically install flows in the flow table of an SDN router for traffic measurement. But there is one constraint: the number of entries in a flow table for traffic measurement is limited because the flow table in an SDN node is usually constructed with expensive ternary content addressable memory (TCAM) devices. How to use the limited number of entries in a flow table for traffic measurement is a challenging task in SDNs. In this paper, we present a new framework for solving the traffic matrix estimation problem in an SDN-based IP network. An important feature of the proposed framework is that any flow added to the flow table of an SDN router can increase the rank of the underlying equations governing the traffic matrix estimation problem. This greatly improves the utilization efficiency of TCAM entries for traffic measurement. Detailed performance evaluation given in this paper demonstrates that the proposed approach can achieve significant performance gains over other approaches.