Biblio

Coupling regular topologies with optimized routing algorithms is key in pushing the performance of interconnection networks of HPC systems. In this paper we present Dmodc, a fast deterministic routing algorithm for Parallel Generalized Fat-Trees (PGFTs) which minimizes congestion risk even under massive topology degradation caused by equipment failure. It applies a modulo-based computation of forwarding tables among switches closer to the destination, using only knowledge of subtrees for pre-modulo division. Dmodc allows complete re-routing of topologies with tens of thousands of nodes in less than a second, which greatly helps centralized fabric management react to faults with high-quality routing tables and no impact to running applications in current and future very large-scale HPC clusters. We compare Dmodc against routing algorithms available in the InfiniBand control software (OpenSM) first for routing execution time to show feasibility at scale, and then for congestion risk under degradation to demonstrate robustness. The latter comparison is done using static analysis of routing tables under random permutation (RP), shift permutation (SP) and all-to-all (A2A) traffic patterns. Results for Dmodc show A2A and RP congestion risks similar under heavy degradation as the most stable algorithms compared, and near-optimal SP congestion risk up to 1% of random degradation.

When discussing data center networks (DCN), topology has a significant influence on the availability of data to the host. The performance of DCN is relative to the scale of the network. On a particular network scale, it can even cause a connection to the host to be disconnected due to the overhead of routing information. It takes a long time to get connected again so that the data packet that has been sent is lost. The length of time for updating routing information to all parts of the topology so that it can be reconnected or referred to as the time of convergence is the cause. Scalability of a network is proportional to the time of convergence. This article discusses Aspen Tree and Fat Tree, which is about the modification of multi-root trees that have been modified. In Fat Tree, a final set of hosts from a network can be disconnected from a network topology until there is an update of routing information that is disseminated to each switch on the network, due to a link failure. Aspen Tree is a reference topology because it is considered to reduce convergence time and control the overhead of network failure recovery. The DCN topology performance models are implemented using the open source NS-3 platform to support validation of performance evaluations.

TCP congestion control has been known for its crucial role in stabilizing the Internet and preventing congestion collapses. However, with the rapid advancement in networking technologies, resulting in the emergence of challenging network environments such as data center networks (DCNs), the traditional TCP algorithm leads to several impairments. The shortcomings of TCP when deployed in DCNs have motivated the development of multiple new variants, including DCTCP, ICTCP, IA-TCP, and D2TCP, but all of these algorithms exhibit their advantages at the cost of a number of drawbacks in the Global Internet. Motivated by the belief that new innovations need to be established on top of a solid foundation with a thorough understanding of the existing, well-established algorithms, we have been working towards a comprehensive analysis of various conventional TCP algorithms in DCNs and other modern networks. This paper presents our first milestone towards the completion of our comparative study in which we present the results obtained by simulating multiple TCP variants: NewReno, Vegas, HighSpeed, Scalable, Westwood+, BIC, CUBIC, and YeAH using a fat tree architecture. Each protocol is evaluated in terms of queue length, number of dropped packets, average packet delay, and aggregate bandwidth as a percentage of the channel bandwidth.