Data centres lie at the heart of almost every service on the Internet. Data centres are used to provide search results, to power social media, to store and index email, to host “cloud” applications, for online retail and to provide a myriad of other web services. Consequently the more efficient they can be made the better for all of us. The power of modern data centres is in combining commodity off-the-shelf server hardware and network equipment to provide what Google’s Barrosso and Ho ̈lzle describe as “warehouse scale” computers. Data centres rely on TCP, a transport protocol that was originally designed for use in the Internet. Like other such protocols, TCP has been optimised to maximise throughput, usually by filling up queues at the bottleneck. However, for most applications within a data centre network latency is more critical than throughput. Consequently the choice of transport protocol becomes a bottleneck for performance. My thesis is that the solution to this is to move away from the use of one-size-fits-all transport protocols towards ones that have been designed to reduce latency across the data centre and which can dynamically respond to the needs of the applications. This dissertation focuses on optimising the transport layer in data centre networks. In particular I address the question of whether any single transport mechanism can be flexible enough to cater to the needs of all data centre traffic. I show that one leading protocol (DCTCP) has been heavily optimised for certain network conditions. I then explore approaches that seek to minimise latency for applications that care about it while still allowing throughput-intensive applications to receive a good level of service. My key contributions to this are Silo and Trevi. Trevi is a novel transport system for storage traffic that utilises fountain coding to max- imise throughput and minimise latency while being agnostic to drop, thus allowing storage traffic to be pushed out of the way when latency sensitive traffic is present in the network. Silo is an admission control system that is designed to give tenants of a multi-tenant data centre guaranteed low latency network performance. Both of these were developed in collaboration with others.

Internet of Things (IoT) is the idea that every device around us is connected and these devices continually collect and communicate data for analysis at a large scale in order to enable better end user experience, resource utilization and device performance. Therefore, data is central to the concept of IoT and the amount being collected is growing at an unprecedented rate. Current networking systems and hardware are not fully equipped to handle influx of data at this scale which is a serious problem because it can lead to erroneous interpretation of the data resulting in low resource utilization and bad end user experience defeating the purpose of IoT. This thesis aims at improving data transportation for IoT. In IoT systems, devices are connected to one or more cloud services over the internet via an access link. The cloud processes the data sent by the devices and sends back appropriate instructions. Hence, the performance of the two ends of the network ie the access networks and datacenter network, directly impacts the performance of IoT. The first portion of the our research targets improvement of the access networks by improving access link (router) design. Among the important design aspects of routers is the size of their output buffer queue. %Selecting an appropriate size of this buffer is crucial because it impacts two key metrics of an IoT system: 1) access link utilization and 2) latency. We have developed a probabilistic model to calculate the size of the output buffer that ensures high link utilization and low latency for packets. We have eliminated limiting assumptions of prior art that do not hold true for IoT. Our results show that for TCP only traffic, buffer size calculated by the state of the art schemes results in at least 60% higher queuing delay compared to our scheme while achieving almost similar access link utilization, loss-rate, and goodput. For UDP only traffic, our scheme achieves at least 91% link utilization with very low queuing delays and aggregate goodput that is approx. 90% of link capacity. Finally, for mixed traffic scenarios our scheme achieves higher link utilization than TCP only and UDP only scenarios as well as low delays, low loss-rates and aggregate goodput that is approx 94% of link capacity. The second portion of the thesis focuses on datacenter networks. Applications that control IoT devices reside here. Performance of these applications is affected by the choice of TCP used for data communication between Virtual Machines (VM). However, cloud users have little to no knowledge about the network between the VMs and hence, lack a systematic method to select a TCP variant. We have focused on characterizing TCP Cubic, Reno, Vegas and DCTCP from the perspective of cloud tenants while treating the network as a black box. We have conducted experiments on the transport layer and the application layer. The observations from our transport layer experiments show TCP Vegas outperforms the other variants in terms of throughput, RTT, and stability. Application layer experiments show that Vegas has the worst response time while all other variants perform similarly. The results also show that different inter-request delay distributions have no effect on the throughput, RTT, or response time. / Master of Science / Internet of Things (IoT) is the idea that every electronic device around us, like watches, thermostats and even refrigerators, is connected to one another and these devices continually collect and communicate data. This data is analyzed at a large scale in order to enable better user experience and improve the utilization and performance of the devices. Therefore, data is central to the concept of IoT and because of the unprecedented increase in the number of connected devices, the amount being collected is growing at an unprecedented rate. Current computer networks over which the data is transported, are not fully equipped to handle influx of data at this scale. This is a serious problem because it can lead to erroneous analysis of the data, resulting in low device utilization and bad user experience, hence, defeating the purpose of IoT. This thesis aims at improving data transportation for IoT by improving different components involved in computer networks. In IoT systems, devices are connected to cloud computing services over the internet through a router. The router acts a gateway to send data to and receive data from the cloud services. The cloud services act as the brain of IoT i.e. they process the data sent by the devices and send back appropriate instructions for the devices to perform. Hence, the performance of the two ends of the network i.e. routers in the access networks and cloud services in datacenter network, directly impacts the performance of IoT. The first portion of our research targets the design of routers. Among the important design aspects of routers is their size of their output buffer queue which holds the data packets to be sent out. We have developed a novel probabilistic model to calculate the size of the output buffer that ensures that the link utilization stays high and the latency of the IoT devices stays low, ensuring good performance. Results show that that our scheme outperforms state-of-the-art schemes for TCP only traffic and shows very favorable results for UDP only and mixed traffic scenarios. The second portion of the thesis focuses on improving application service performance in datacenter networks. Applications that control IoT devices reside in the cloud and their performance is directly affected by the protocol chosen to send data between different machines. However, cloud users have almost no knowledge about the configuration of the network between the machines allotted to them in the cloud. Hence, they lack a systematic method to select a protocol variant that is suitable for their application. We have focused on characterizing different protocols: TCP Cubic, Reno, Vegas and DCTCP from the perspective of cloud tenants while treating the network as a black-box (unknown). We have provided in depth analysis and insights into the throughput and latency behaviors which should help the cloud tenants make a more informed choice of TCP congestion control.